<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Luiz Figueiredo on Medium]]></title>
        <description><![CDATA[Stories by Luiz Figueiredo on Medium]]></description>
        <link>https://medium.com/@figueiredoluiz?source=rss-67f54c9d481d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*wz8fkJBlBBlTWs3Va44IwA.png</url>
            <title>Stories by Luiz Figueiredo on Medium</title>
            <link>https://medium.com/@figueiredoluiz?source=rss-67f54c9d481d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 08:40:19 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@figueiredoluiz/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[AI Made Code Cheap. It Didn’t Make Software Engineering Obsolete.]]></title>
            <link>https://figueiredoluiz.medium.com/ai-made-code-cheap-it-didnt-make-software-engineering-obsolete-243e7ef19128?source=rss-67f54c9d481d------2</link>
            <guid isPermaLink="false">https://medium.com/p/243e7ef19128</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Luiz Figueiredo]]></dc:creator>
            <pubDate>Sat, 21 Nov 2026 07:10:02 GMT</pubDate>
            <atom:updated>2026-11-21T07:23:18.389Z</atom:updated>
            <content:encoded><![CDATA[<p><em>AI is accelerating implementation. The harder challenge is making verification, integration, and reliability scale with it.</em></p><figure><img alt="Abstract code streams funneling through a narrow verification gate before becoming shipped software." src="https://cdn-images-1.medium.com/max/1024/1*P5PisHHUxR81k0ot3IJRqw.png" /></figure><p>I use AI coding tools extensively. If an agent can produce a correct implementation in minutes, I see no reason to spend hours writing it myself. I’d rather use that time to understand the problem, explore alternatives, or improve the system around the implementation.</p><p>What I question is the assumption that cheaper code makes software engineering less relevant. Engineering has always involved requirements, architecture, verification, maintenance, and decisions about acceptable risk. Even writing a prompt involves engineering work: someone defines the desired behavior, establishes constraints, and decides what a satisfactory result should look like.</p><p>An experienced engineer and a newcomer can use the same model and still build fundamentally different systems. Their differences may be invisible during a demonstration and emerge only when real users, failures, changing requirements, and maintenance enter the picture.</p><p>Making software accessible to more people, including through <em>vibe coding</em>, is a remarkable achievement. But a working prototype and a dependable product have different requirements, and the appropriate level of engineering depends on what we’re building and the consequences of failure.</p><p><strong>Making implementation cheap doesn’t make engineering decisions free.</strong></p><p>Recent research helps examine that distinction. One study estimated a 240% increase in commits associated with autonomous coding agents, but only a 30% increase in releases [4]. Both figures are positive, yet they describe different stages of production.</p><p>The gap directs attention to what happens after code generation. Our capacity to produce changes may grow faster than our capacity to evaluate them.</p><h3>1. Code production is accelerating. But what are we measuring?</h3><p>In 2026, <em>Management Science</em> published research combining three randomized field experiments involving <strong>4,867 developers</strong> at Microsoft, Accenture, and an unnamed Fortune 100 company [1]. Developers given access to AI coding assistance completed an estimated <strong>26.08% more tasks</strong>, with generally larger gains among less-experienced participants.</p><p>These controlled experiments provide valuable evidence about productivity in real working environments. However, the researchers measured development activity, including pull-request-based outcomes, rather than directly measuring product value or total engineering cost.</p><p>A developer submitting three PRs isn’t necessarily delivering three times the value of someone submitting one. The researchers weren’t claiming otherwise, but the distinction can disappear when carefully defined experimental results become broad productivity claims.</p><p>An earlier METR experiment pointed in a different direction. In a randomized study involving 16 experienced open-source developers completing 246 tasks in familiar repositories, access to early-2025 AI tools made participants approximately <strong>19% slower</strong> [2]. The developers had expected a speedup and even believed AI had helped afterward.</p><p>The findings need to be read in context. These studies examined different populations, tasks, environments, and productivity measures. METR’s smaller experiment also has different limits on generalization. In February 2026, METR reported preliminary results with newer tools that suggested possible improvements, while emphasizing recruitment and measurement problems that prevented a reliable new estimate [3].</p><p>The technology and the way developers use it are evolving rapidly. That makes universal productivity claims difficult to sustain, even when individual studies provide strong evidence for their specific settings.</p><h3>Writing code versus shipping code</h3><p>Another perspective comes from <em>Writing Code vs. Shipping Code</em>, a 2026 NBER working paper by Mert Demirer, Leon Musolff, and Liyuan Yang [4]. Demirer and Musolff also participated in the earlier study reporting the 26.08% productivity gain.</p><p>Using data from more than <strong>500,000 GitHub developers</strong> and AI usage telemetry, the researchers examined successive generations of coding tools. In the September 2026 revision, their matched observational analysis estimated cumulative increases in commits of <strong>30% for autocomplete, 180% for interactive agents, and 240% for autonomous agents</strong>.</p><p>The effects became smaller farther along the production process. With autonomous agents, the estimated increase was <strong>80% in projects and 30% in releases</strong>. Releases did increase, but implementation activity grew much faster than downstream output.</p><figure><img alt="Two bar charts showing estimated commit increases by AI tool generation, and smaller downstream increases in projects and releases for autonomous agents." src="https://cdn-images-1.medium.com/max/1024/1*iKqTZRDtBPloqycDy4L-Ow.png" /><figcaption><em>Figure 1. Activity increased across all measured outcomes, but the estimated gain was smaller at later stages. Observational NBER results, not a causal breakdown of delivery time [4].</em></figcaption></figure><p>The authors also examined four software marketplaces and found a substantial increase in new applications without a corresponding increase in aggregate usage. Generating code, shipping software, and creating something people actually use are distinct outcomes.</p><p>We can’t attribute the entire difference to review bottlenecks. Cheaper implementation also makes experimentation cheaper: teams can discard branches, explore alternatives, and expand feature scope. Some additional code may represent productive exploration rather than waste.</p><p>The researchers interpret these diminishing downstream effects as consistent with a <em>weak-link production</em> model. When an outcome depends on several activities, speeding up one doesn’t necessarily accelerate the result proportionally. The paper also estimates strong complementarity between AI output and human effort, rather than treating them as perfect substitutes.</p><p>What I find most relevant is the distinction between <strong>implementation capacity and delivery capacity</strong>. Code review offers a setting to investigate why these capacities don’t always grow together.</p><h3>2. Cheap contributions can create expensive decisions</h3><p>When an agent submits a change, someone — or something — must establish whether it should exist, whether it behaves correctly, and whether it can be integrated safely. Generating a plausible implementation may take minutes, while evaluating its consequences can require far more context.</p><p>A November 2026 study by Newman and colleagues combined 24 interviews with a survey of <strong>2,590 developers</strong> at a large multinational company [5]. Participants described larger pull requests, increased review volume, and cases where authors couldn’t adequately explain AI-assisted changes.</p><p>Human reviewers were increasingly concerned with intent, design, and architecture. Meanwhile, using AI for review was <strong>not correlated with less difficulty keeping up with PR volume</strong>.</p><p>The authorship problem deserves attention. Traditionally, we expect someone submitting code to understand the implementation well enough to explain it and respond to feedback. With coding agents, that understanding can no longer be assumed. A reviewer may inherit investigative work that should have happened before submission.</p><p>Adding another agent to review the first agent’s output may help, but its effectiveness still depends on the criteria used to decide whether a change is acceptable.</p><h3>Passing tests isn’t the same as being ready to merge</h3><p>METR examined this distinction by asking maintainers to assess AI-generated patches from SWE-bench Verified [6]. Four maintainers across three repositories evaluated 296 patches, with previously merged human patches used to account for noise and disagreement in review decisions.</p><p>After adjustment, the automated grader overstated maintainer acceptance by <strong>about 24 percentage points</strong>. Roughly half of the test-passing patches would not have met the maintainers’ acceptance standards.</p><figure><img alt="Two separate METR findings: roughly half of test-passing AI patches fail maintainer standards, and the adjusted acceptance-rate overstatement is about 24 percentage points." src="https://cdn-images-1.medium.com/max/1024/1*LJHMWmGCyF5UcSqkklNCVA.png" /><figcaption><em>Figure 2. A test-passing patch may still fail maintainer review. The illustrative half-and-half split and adjusted 24-point acceptance gap are distinct summaries [6].</em></figcaption></figure><p>The experiment didn’t allow agents to iterate on review feedback as human contributors typically would. That limits what we can conclude about their capabilities, but the result still illustrates the difference between passing an automated grader and satisfying a project’s broader acceptance criteria.</p><p>A test suite verifies the behaviors it was designed to check. It cannot guarantee that every relevant behavior was considered.</p><p>This question also motivates some of my work on <a href="https://vitestdedup.dev/">VitestDedup</a> and <a href="https://github.com/figueiredoluiz/cuke-dedup">CukeDedup</a>. VitestDedup focuses on duplicate and near-duplicate Vitest tests, while CukeDedup analyzes duplication and ambiguity in Cucumber/Gherkin step definitions.</p><p>Both address problems that existed long before AI, but generating larger volumes of tests and test infrastructure makes those problems increasingly relevant. An agent may produce dozens of tests in a single iteration, yet some may repeat the same expectations with different inputs rather than verify distinct behaviors.</p><p>More passing tests can be valuable, but test count alone doesn’t establish greater defect-detection capability. Evaluating the quality of generated tests deserves as much attention as generating them.</p><h3>Open source makes the problem visible</h3><p>Open-source maintainers face a difficult version of this asymmetry. Contributors can generate and submit changes cheaply, while maintainers bear much of the cost of deciding whether those changes deserve integration.</p><p>Afroz and colleagues examined <strong>294 repositories and more than two million pull requests and issues</strong>, finding increased contribution volume alongside declining merge rates [7]. For one-time contributors, merge rates were estimated to be 18.18% below the modeled counterfactual.</p><p>The Linux Foundation separately reported productivity benefits from AI alongside growing pressure on maintainers. In its 2026 survey, respondents estimated that approximately 25% of incoming PRs were low-value AI-associated contributions [8]. That percentage reflects maintainers’ judgments, not an independent classification of all AI-generated pull requests.</p><p>Mem0 provides a concrete example. In August 2026, a maintainer reported <strong>462 open pull requests</strong>, roughly <strong>100 new submissions per week</strong>, and around <strong>30 weekly merges</strong> [9]. The project subsequently implemented a contribution gate requiring external PRs to reference an issue marked accepted by a maintainer. PRs without an accepted issue are automatically closed, while documentation-only changes are exempt.</p><p>Its contribution guide requires authors to understand and explain their changes, including how they interact with the rest of the codebase, without the help of an AI tool. AI-generated implementations and descriptions are permitted, but contributors remain responsible for understanding what they submit.</p><p>Mem0’s approach illustrates a practical way to allocate limited attention. Maintainers first establish whether the underlying problem deserves a change, rather than reviewing every submitted implementation.</p><p>GitHub also acknowledged the growing pressure in February 2026, describing an <em>Eternal September</em> for open source [10]. In June, it introduced configurable limits on open pull requests from users without write access [11].</p><p>These examples document a capacity problem in parts of the open-source ecosystem. When contributions arrive faster than maintainers can evaluate them, legitimate work becomes harder to process. The response should focus on identifying valuable changes and improving evaluation capacity, regardless of whether AI helped produce them.</p><h3>3. The answer isn’t more reviewers. It’s better verification.</h3><p>If AI can multiply implementation output, increasing manual review proportionally would be an expensive way to scale development. I’d rather ask which decisions can be automated reliably and which deserve deeper human attention.</p><p>Meta’s <strong>RADAR</strong> offers a concrete example.</p><p>In <em>Automating Low-Risk Code Review at Meta</em>, researchers describe a production system designed to evaluate and integrate selected low-risk changes automatically [12].</p><p>At Meta, significant lines of code per landed human diff increased <strong>105.9% year over year</strong>, while diff volume per developer rose <strong>51%</strong>. According to the paper, agentic AI accounted for more than 80% of that volume growth. Meanwhile, the proportion of changes receiving timely review was declining.</p><p>Meta responded by redesigning part of its verification pipeline. RADAR combines origin classification, eligibility checks, static heuristics, a machine-learned risk score, LLM-based review, and deterministic validation before qualifying changes can be integrated automatically.</p><p>The study reports more than <strong>535,000 diffs reviewed</strong> through RADAR and over <strong>331,000 integrated</strong>. When the risk eligibility threshold expanded from the 25th to the 50th percentile, the approval rate reached <strong>60.31%</strong>. Researchers also reported a <strong>35% reduction in median review wall time</strong>.</p><p>The observed revert rate among RADAR-reviewed changes was one-third that of non-RADAR changes, while the observed production-incident rate was one-fiftieth. These figures require careful interpretation because RADAR deliberately selects lower-risk changes. The comparison doesn’t establish that automated review caused improvements of those magnitudes relative to human review.</p><p>Even with that limitation, the operational lesson holds: Meta adapted verification to rising code volume while preserving stronger controls for changes that didn’t qualify for automatic integration.</p><figure><img alt="Simplified RADAR review pipeline: source and origin classification, eligibility, static heuristics, ML risk scoring, LLM review, deterministic checks, then automatic integration for eligible changes or another review path otherwise." src="https://cdn-images-1.medium.com/max/1024/1*mXYHA2t9YPzFp2qvCXzq-g.png" /><figcaption><em>Figure 3. Simplified RADAR review pipeline and selected operating metrics. The lower observed incident and revert rates reflect a low-risk selected population; causal superiority over human review is not established [12].</em></figcaption></figure><h3>What would that mean for an ordinary repository?</h3><p>Most teams don’t have Meta’s infrastructure, but the principle is transferable. A documentation-only change might require formatting and link checks, with little manual attention. A conventional application change could pass through automated tests, static analysis, and review appropriate to its impact.</p><p>Changes affecting authentication, financial calculations, destructive migrations, or critical data contracts deserve stronger verification and explicitly responsible reviewers. The precise requirements should reflect the consequences of failure rather than the amount of code generated. These examples illustrate risk-based allocation of verification effort, not universal rules validated by the RADAR experiment.</p><figure><img alt="Illustrative risk-based verification tiers: documentation-only changes receive light automated checks, ordinary application changes receive tests and proportionate review, and high-impact changes receive stronger checks with accountable human reviewers." src="https://cdn-images-1.medium.com/max/1024/1*ZS9Stj1kwkm5hbTQ264X5A.png" /><figcaption><em>Figure 4. An illustrative policy for matching verification effort to the consequences of failure, rather than code volume. Not validated as a universal RADAR policy [12].</em></figcaption></figure><p>The same principle applies before a pull request exists. Teams can define acceptance criteria, establish important invariants, and decide what evidence would demonstrate that an implementation is safe to integrate.</p><p>Agents can help with those activities as well as implementation. Using AI to generate tests, review changes, and enforce architectural rules is compatible with rigorous software engineering.</p><p><strong>The goal isn’t to preserve manual work. It’s to make automation trustworthy.</strong></p><h3>4. When trust becomes a throughput constraint</h3><p>The available evidence doesn’t support a universal verdict on whether AI improves or worsens code quality. The differences between studies are more informative than a single answer.</p><p>DORA’s 2025 report described AI as an amplifier of existing organizational strengths and weaknesses, emphasizing the importance of engineering foundations [13]. Faros AI subsequently reported a more troubling picture.</p><p>Its Q2 2026 <em>Acceleration Whiplash</em> report, based on engineering telemetry from approximately <strong>22,000 developers across 4,000 teams</strong>, associated higher AI adoption with larger PRs, longer reviews, greater rework, and more incidents [14].</p><p>The Q3 report, <em>The Speed Trap</em>, adds a temporal contrast [15]. Faros reported that incidents per PR were <strong>14.5% higher</strong> under high AI adoption in the newer analysis, compared with a <strong>242.7% increase</strong> in the earlier report. The newer report also found that time in QA was up <strong>300.6%</strong>, PRs merged without review were up <strong>76.3%</strong>, and monthly incidents were up <strong>125.4%</strong>.</p><p>These figures come from observational comparisons across reporting periods and analytical conditions. They don’t establish causal improvements attributable to AI, nor should they be read as results from a single controlled longitudinal experiment.</p><figure><img alt="Faros AI reported associations across Q2 and Q3 2026: incidents per PR up 242.7% in Q2 and 14.5% in Q3, plus Q3 time in QA up 300.6%, monthly incidents up 125.4%, and unreviewed merges up 76.3%." src="https://cdn-images-1.medium.com/max/1024/1*PPMlA8AcoFC3oyO51kgnMA.png" /><figcaption><em>Figure 5. Reported associations from two Faros AI studies. Each percentage refers to that report’s own comparison, not a controlled before-and-after experiment [14][15].</em></figcaption></figure><p>They raise a question about aggregate operational burden. Even if problems associated with individual changes become less frequent, a sufficiently large increase in change volume can produce more incidents and more work downstream.</p><p>DORA and Faros disagree about how effectively mature engineering practices protect organizations during rapid AI adoption. Their methods and definitions of maturity differ. Faros sells engineering observability software, so its commercial context matters when interpreting its conclusions.</p><p>A separate large-scale study at Meta further illustrates the importance of context [16]. Across more than <strong>10,000 developers and 400,000 diffs</strong>, AI-assisted changes were associated with faster reviews and fewer severe incidents. But AI use was concentrated in less-central code, larger changes, test files, and work by more-tenured authors.</p><p>These differences complicate attempts to attribute favorable outcomes directly to AI. They also suggest that the effectiveness of engineering controls may depend on where, how, and by whom the technology is used.</p><p>One possible feedback loop is that teams accelerate implementation while escaped defects and rework accumulate. Confidence in the delivery process may then decline, leading to more approvals, longer QA cycles, and slower deployments that consume some of the time saved during implementation.</p><p>A rollback isn’t necessarily evidence of poor engineering; rapid detection and recovery can indicate a healthy delivery process. What matters is whether a team can continue changing software without allowing the cost of mistakes to overwhelm the benefits of speed.</p><p><strong>Reliable verification is part of what makes sustained productivity possible.</strong></p><h3>Engineering is what makes acceleration sustainable</h3><p>I don’t believe engineers need to defend the act of writing code. If AI can reliably eliminate manual work, that’s progress. It may allow smaller teams to build more sophisticated products, reduce costs, or make previously uneconomical projects viable.</p><p>The research discussed here doesn’t establish the net effect on engineering employment. It does challenge the assumption that automating implementation makes engineering judgment, verification, and accountability unnecessary.</p><p>Some existing processes deserve to disappear as well. Reviews can be redundant, tests can provide little protection, and approval chains sometimes reflect organizational history rather than technical risk. We should improve or eliminate those mechanisms, using automation where appropriate.</p><p>RADAR demonstrates a direction I find promising: engineering can make automated review itself more scalable. The discipline extends beyond performing the work to designing the systems that perform and verify it.</p><p><strong>AI accelerates software production. Engineering turns that acceleration into dependable outcomes.</strong></p><p>As agents take on more implementation, testing, review, and operations, our ability to define acceptable outcomes matters more, not less.</p><p>One question remains: as we delegate more work to agents, what happens to our ability to recognize when they’re wrong?</p><p>In January 2026, Anthropic published a randomized experiment involving <strong>52 mostly junior developers</strong> learning a Python library [17]. Participants using AI assistance scored an average of <strong>50%</strong> on a subsequent comprehension assessment, compared with <strong>67%</strong> among those working without AI.</p><p>The study examined immediate learning in a specific task, not permanent cognitive deterioration. It also found differences in learning outcomes depending on how participants used AI.</p><p>We can build increasingly sophisticated mechanisms to verify generated code, but confidence in those mechanisms still depends on understanding what they check, what they miss, and when their conclusions deserve scrutiny.</p><p><em>Perhaps the hardest risk to measure won’t be the code AI writes, but the understanding we stop exercising while it writes.</em></p><h3>References</h3><p><strong>[1]</strong> Cui, K. Z., et al. (2026). <em>The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers.</em> Management Science.<br><a href="https://doi.org/10.1287/mnsc.2025.00535">https://doi.org/10.1287/mnsc.2025.00535</a></p><p><strong>[2]</strong> Becker, J., et al. (2025). <em>Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.</em> METR.<br><a href="https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/">https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/</a></p><p><strong>[3]</strong> METR (2026). <em>We are Changing our Developer Productivity Experiment Design.</em><br><a href="https://metr.org/blog/2026-02-24-uplift-update/">https://metr.org/blog/2026-02-24-uplift-update/</a></p><p><strong>[4]</strong> Demirer, M., Musolff, L., &amp; Yang, L. (2026). <em>Writing Code vs. Shipping Code: Productivity Effects Across Generations of AI Coding Tools.</em> NBER Working Paper 35275, September 2026 revision.<br><a href="https://www.nber.org/papers/w35275">https://www.nber.org/papers/w35275</a></p><p><strong>[5]</strong> Newman, K., Bird, C., Butler, J., &amp; Murphy-Hill, E. (2026). <em>Level Up! AI-Driven Challenges and Adaptations to Code Review Practices.</em> Preprint.<br><a href="https://arxiv.org/abs/2611.18939">https://arxiv.org/abs/2611.18939</a></p><p><strong>[6]</strong> Whitfill, P., Wu, C., Becker, J., &amp; Rush, N. (2026). <em>Many SWE-bench-Passing PRs Would Not Be Merged into Main.</em> METR.<br><a href="https://metr.org/notes/2026-03-10-many-swe-bench-passing-prs-would-not-be-merged-into-main/">https://metr.org/notes/2026-03-10-many-swe-bench-passing-prs-would-not-be-merged-into-main/</a></p><p><strong>[7]</strong> Afroz, S., et al. (2026). <em>“AI Slop is DDoSing Open Source”: Understanding the Impact of AI-Generated Contributions on Open Source Sustainability.</em> Preprint.<br><a href="https://arxiv.org/abs/2607.04003">https://arxiv.org/abs/2607.04003</a></p><p><strong>[8]</strong> Linux Foundation Research (2026). <em>Navigating Open Source Software Development in the Generative AI Era.</em><br><a href="https://www.linuxfoundation.org/press/new-lf-research-survey-shows-generative-ai-cuts-open-source-coding-time-in-half-but-25-of-incoming-pull-requests-are-considered-ai-slop">https://www.linuxfoundation.org/press/new-lf-research-survey-shows-generative-ai-cuts-open-source-coding-time-in-half-but-25-of-incoming-pull-requests-are-considered-ai-slop</a></p><p><strong>[9]</strong> Mem0 (2026). <em>Gate pull requests on an accepted issue to stop unverified AI-generated contributions.</em> Issue #6893.<br><a href="https://github.com/mem0ai/mem0/issues/6893">https://github.com/mem0ai/mem0/issues/6893</a><br>Implemented policy: <a href="https://github.com/mem0ai/mem0/blob/main/CONTRIBUTING.md">https://github.com/mem0ai/mem0/blob/main/CONTRIBUTING.md</a></p><p><strong>[10]</strong> GitHub (2026). <em>Welcome to the Eternal September of open source.</em><br><a href="https://github.blog/open-source/maintainers/welcome-to-the-eternal-september-of-open-source-heres-what-we-plan-to-do-for-maintainers/">https://github.blog/open-source/maintainers/welcome-to-the-eternal-september-of-open-source-heres-what-we-plan-to-do-for-maintainers/</a></p><p><strong>[11]</strong> GitHub Changelog (June 17, 2026). <em>Limit open pull requests for users without write access.</em><br><a href="https://github.blog/changelog/2026-06-17-limit-open-pull-requests-for-users-without-write-access/">https://github.blog/changelog/2026-06-17-limit-open-pull-requests-for-users-without-write-access/</a></p><p><strong>[12]</strong> Adams, C., et al. (2026). <em>Automating Low-Risk Code Review at Meta: RADAR, Risk Calibration, and Review Efficiency.</em><br><a href="https://arxiv.org/abs/2605.30208">https://arxiv.org/abs/2605.30208</a></p><p><strong>[13]</strong> DORA / Google Cloud (2025). <em>State of AI-assisted Software Development.</em><br><a href="https://dora.dev/research/2025/dora-report/">https://dora.dev/research/2025/dora-report/</a></p><p><strong>[14]</strong> Faros AI (Q2 2026). <em>The Acceleration Whiplash.</em><br><a href="https://www.faros.ai/research/ai-acceleration-whiplash">https://www.faros.ai/research/ai-acceleration-whiplash</a></p><p><strong>[15]</strong> Faros AI (Q3 2026). <em>The Speed Trap.</em><br><a href="https://www.faros.ai/research/speed-trap">https://www.faros.ai/research/speed-trap</a></p><p><strong>[16]</strong> Mockus, A., et al. (2026). <em>A Preliminary Analysis of the Impact of AI Assisted/Generated Code on Quality, Centrality and Review Time at Scale.</em> ESEM 2026.<br><a href="https://doi.org/10.4230/LIPIcs.ESEM.2026.93">https://doi.org/10.4230/LIPIcs.ESEM.2026.93</a></p><p><strong>[17]</strong> Anthropic Research (2026). <em>How AI Assistance Impacts the Formation of Coding Skills.</em><br><a href="https://www.anthropic.com/research/AI-assistance-coding-skills">https://www.anthropic.com/research/AI-assistance-coding-skills</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=243e7ef19128" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[October 2026, the Witching Month of Patches]]></title>
            <link>https://figueiredoluiz.medium.com/october-2026-the-witching-month-of-patches-1234898a6adb?source=rss-67f54c9d481d------2</link>
            <guid isPermaLink="false">https://medium.com/p/1234898a6adb</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[infosec]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <dc:creator><![CDATA[Luiz Figueiredo]]></dc:creator>
            <pubDate>Sun, 01 Nov 2026 02:38:21 GMT</pubDate>
            <atom:updated>2026-11-01T02:38:21.221Z</atom:updated>
            <content:encoded><![CDATA[<h4>Are we being haunted, or did someone finally turn the lights on?</h4><figure><img alt="Cover illustration: a flashlight beam reveals ghostly software bugs around a haunted house." src="https://cdn-images-1.medium.com/max/1024/1*dp3OTAt4pdUb1ZAtYzL_mQ.png" /></figure><h3>Trick or treat, update your browser</h3><p>If your laptop felt clingy this week, you were not imagining it. On Monday, October 26, CISA <a href="https://www.cisa.gov/news-events/alerts/2026/10/26/cisa-adds-two-known-exploited-vulnerabilities-catalog">added two PHP flaws</a> — the “PHP2Shell” pair — to its Known Exploited Vulnerabilities catalog. On Tuesday, Firefox 159 shipped with <a href="https://www.mozilla.org/en-US/security/advisories/mfsa2026-118/">a security advisory containing 167 CVEs</a>. On Wednesday, WordPress 7.1.4 <a href="https://wordpress.org/documentation/wordpress-version/version-7-1-4/">addressed 22 vulnerabilities</a>, its second security release of the month. That same week, Google published a <a href="https://chromereleases.googleblog.com/2026/10/stable-channel-update-for-desktop_0562499947.html">Chrome 156 update containing 52 security fixes</a>. Earlier in October, independent tallies of Microsoft’s Patch Tuesday exceeded 1,400 vulnerabilities, and Oracle’s quarterly update delivered <a href="https://www.oracle.com/security-alerts/cpuoct2026.html">1,295 new security patches</a>.</p><p>For anyone outside security, the headline writes itself: software is falling apart. My hypothesis is the opposite — or, at least, more boring. Many of these monsters were already living in the house. What changed in 2026 is that we got much better flashlights.</p><p>That is a hypothesis, not a conclusion. So I collected historical numbers and accounts from the teams shipping the fixes to see where it holds and where it breaks.</p><h3>How big is the haunting?</h3><figure><img alt="Bar chart: Published CVE records for January–September of 2023, 2024, 2025, and 2026." src="https://cdn-images-1.medium.com/max/1024/1*LnlHoGbC5UE4UCd5_sbn6w.png" /></figure><p>First, an apples-to-apples comparison using official quarterly metrics. These are <em>published CVE records</em>, not the dates on which flaws entered the codebase.</p><p><em>Source: </em><a href="https://www.cve.org/About/Metrics"><em>CVE Program — published CVE records by quarter</em></a><em>. All bars cover January through September, not full calendar years.</em></p><p>For context, the same official source lists full-year totals of 28,961 for 2023, 40,077 for 2024, and 48,244 for 2025. By September 30, the 2026 count had already passed all three. The third quarter of 2026 alone accounted for 36,962 published records, compared with 11,738 in Q3 2025 — about 3.1 times as many.</p><p>The growth did not start in 2026, but 2026 is a different animal. In June, <a href="https://www.first.org/blog/20260615-vulnerability-forecast-update">FIRST’s forecasting team revised its projection</a> to approximately 66,000 CVEs for the entire year. The published total exceeded that forecast by the end of September. That does not mean forecasting failed in a simple sense: the team had explicitly discussed unusually wide uncertainty ranges and the possibility of AI-driven discovery exceeding expectations.</p><p>Now narrow the lens to October. Not every vendor uses the same unit, so these examples should not be added together.</p><figure><img alt="Table: October 2025 vs. October 2026 for Oracle Critical Patch Update (374 vs. 1,295 new patches), WordPress Core (0 vs. 2 security releases) and Microsoft Patch Tuesday (172 vs. 1,411, independent media counts)." src="https://cdn-images-1.medium.com/max/1024/1*ZUxYMNrdFx_kcuCSLGNJRg.png" /></figure><p><em>Sources by row: Oracle’s </em><a href="https://www.oracle.com/security-alerts/cpuoct2025.html"><em>October 2025</em></a><em> and </em><a href="https://www.oracle.com/security-alerts/cpuoct2026.html"><em>October 2026</em></a><em> advisories; WordPress’s </em><a href="https://wordpress.org/documentation/wordpress-version/version-7-1-3/"><em>7.1.3</em></a><em> and </em><a href="https://wordpress.org/documentation/wordpress-version/version-7-1-4/"><em>7.1.4</em></a><em> release notes; BleepingComputer’s independently compiled </em><a href="https://www.bleepingcomputer.com/news/microsoft/microsoft-october-2025-patch-tuesday-fixes-6-zero-days-172-flaws/"><em>Microsoft October 2025</em></a><em> and </em><a href="https://www.bleepingcomputer.com/news/microsoft/microsoft-october-2026-patch-tuesday-fixes-1-411-flaws-3-zero-days/"><em>October 2026</em></a><em> counts. The Microsoft numbers are </em><strong><em>not</em></strong><em> official standardized totals; the </em><a href="https://msrc.microsoft.com/update-guide/"><em>MSRC Update Guide</em></a><em> is the canonical source for a future like-for-like recount.</em></p><p>Even that WordPress contrast is sensitive to the calendar: its <a href="https://wordpress.org/news/2025/09/wordpress-6-8-3-security-release/">6.8.3 security release</a> landed on September 30, 2025, just one day before the comparison window. A monthly tally tells us when fixes shipped, not how secure the project was the day before.</p><p>So yes, October 2026 looks exceptional across several major release programs. But the numbers alone cannot tell us what kind of exception it is.</p><h3>Where the ghosts come from</h3><figure><img alt="Bar chart: Firefox security bugs fixed per month, 20 to 30 in a typical month of 2025 versus 423 in April 2026, of which 271 were attributed to Claude Mythos Preview." src="https://cdn-images-1.medium.com/max/1024/1*3uYXKFuGaLxv0vWF0bpjAw.png" /></figure><p>The strongest evidence for the flashlight hypothesis does not come from CVE totals. It comes from maintainers describing their own workloads.</p><p>Mozilla was fixing roughly 20 to 30 security bugs per month in Firefox throughout 2025. In April 2026 that number was <strong>423</strong>. The team attributed <strong>271</strong> of those fixes to bugs identified with Anthropic’s Claude Mythos Preview. In a detailed <a href="https://hacks.mozilla.org/2026/05/behind-the-scenes-hardening-firefox/">Mozilla Hacks post</a>, Firefox engineers explained how they built an agentic discovery pipeline on top of their existing fuzzing infrastructure.</p><p>Their account is more interesting than a simple “AI found the bugs” headline. The gains came from better models <em>and</em> better engineering: reproducible test cases, ephemeral virtual machines, deduplication, triage, and integration with Firefox’s existing fix-and-release workflow. The key point is not merely that the models were stronger; it is that the surrounding harness and validation systems made their findings operationally useful.</p><p>On October 29, Firefox’s Bobby Holley <a href="https://blog.mozilla.org/en/firefox/privacy-security/defending-patch-gap/">reported</a> that the team had shipped <strong>more than seven times as many security fixes</strong> as during the corresponding period of 2025, using early access to frontier models from Anthropic and OpenAI. The <a href="https://www.mozilla.org/en-US/security/advisories/mfsa2026-118/">Firefox 159 advisory</a> also names OpenAI researchers and OpenAI Codex Security among the reporters.</p><figure><img alt="Bar chart: WordPress HackerOne reports per month, 20 to 30 in a typical month of the previous decade versus 450 in July 2026 and 773 in August 2026." src="https://cdn-images-1.medium.com/max/1024/1*93gH_dq8jhSJLA92naMZvw.png" /></figure><p>WordPress tells the same story from the inbox side. Its HackerOne program had received 20 to 30 reports per month for roughly a decade. In July 2026 it received 450; in August, 773. According to <a href="https://make.wordpress.org/security/2026/10/29/the-core-security-initiative-two-months-in/">Rudy Faile’s October 29 report</a>, the project had received more than 3,000 reports in 2026 — more than the previous twelve years combined, excluding spam and reports marked not applicable.</p><p>The team has responded with substantial changes, but it would be misleading to say the backlog has simply disappeared. WordPress reports seven Core security releases since mid-July, closing more than 50 critical or high-severity issues, with fixes backported to older supported branches as far back as 4.7. It also automated approximately 73% of the 109 steps in its security-release process. That is an operational response to the increased volume, not proof that maintainers can absorb every report without delay.</p><p>GitHub saw the wave early. In May 2026, its Advisory Database <a href="https://github.blog/security/supply-chain-security/inside-the-advisory-database-and-what-happens-when-vulnerability-volume-breaks-records/">published 1,560 reviewed advisories</a>, more than five times its typical monthly output. Private vulnerability reports went from about 550 per week in January to more than 3,000 per week for much of May, while GitHub’s CVE Numbering Authority received nearly 4,000 CVE requests in May alone. Review times grew from days to weeks for a meaningful share of submissions.</p><p>The bottleneck did not disappear when discovery got cheaper; it shifted toward validating, deduplicating, coordinating, fixing, and publishing the findings.</p><p>Independent <a href="https://epoch.ai/data-insights/cve-severity-spike">analysis from Epoch AI</a> likewise found a dramatic spike in high- and critical-severity CVE disclosures among major organizations in mid-2026. The timing is consistent with the emergence of more capable security research models, but timing alone is not proof of causation.</p><p>This matters for the thesis. We have directly documented examples of old, previously undiscovered bugs being found faster. We <strong>do not</strong> have sufficient data to say that almost all 2026 vulnerabilities were old, or that newly written code has become better or worse. A CVE published in October 2026 can describe a bug that shipped years earlier; publication date is not introduction date.</p><h3>Part of the monster is the measuring tape</h3><p>Before celebrating, there is an uncomfortable confound: the way we count changed too.</p><p>Mozilla’s <a href="https://www.mozilla.org/en-US/security/advisories/mfsa2026-118/">Firefox 159 advisory</a> notes that internally found memory-safety bugs are now disclosed as <strong>individual CVEs</strong> instead of being bundled under a single umbrella identifier. That makes the headline number bigger even when the underlying engineering work is similar.</p><p>The broader CVE system is changing as well. The <a href="https://www.cve.org/About/Metrics">CVE Program’s metrics page</a> shows that the number of active CVE Numbering Authorities increased from <strong>263 at the end of 2022</strong> to <strong>563 by the end of Q3 2026</strong>. Meanwhile, the share of published records issued directly by CNAs rose from <strong>74% in 2022</strong> to <strong>95% by Q3 2026</strong>. More publishers mean more records, regardless of whether the software itself got any less safe.</p><p>Even Microsoft’s October Patch Tuesday demonstrates how counting methodology matters. Different analysts produced different totals because they may include or exclude previously disclosed issues, third-party component CVEs, reissued updates, or other categories. For a consistent multi-year comparison, we would need to reprocess the <a href="https://msrc.microsoft.com/update-guide/">MSRC data</a> under one explicit set of inclusion rules.</p><p>The infrastructure is under strain. In September, CISA published a <a href="https://www.cisa.gov/sites/default/files/2026-09/cve-program-establishing-a-quality-era-framework-508c.pdf">“Quality Era” framework for the CVE Program</a>, reflecting concern about maintaining useful records as disclosure volume grows. GitHub has similarly described the importance of high-quality version ranges and reproducible evidence when every curation bottleneck compounds downstream.</p><p>The curve is real. But it is a measure of <em>published records</em>, not a direct readout of how unsafe today’s software is.</p><h3>The real test: are more of these actually being exploited?</h3><figure><img alt="Bar chart: CISA Known Exploited Vulnerabilities additions by year: 187 in 2023, 186 in 2024, 245 in 2025, and 246 in January–September 2026." src="https://cdn-images-1.medium.com/max/1024/1*bvkB-tMJbk1tSS5TyZVAbA.png" /></figure><p>If growing disclosure volumes reflected a proportionate growth in actual attacks, we would expect stronger supporting evidence from exploitation data. One useful — though incomplete — source is <a href="https://github.com/cisagov/kev-data">CISA’s Known Exploited Vulnerabilities (KEV) catalog</a>, which includes vulnerabilities for which CISA has evidence of exploitation in the wild.</p><p>For this article, I counted entries by their <strong>dateAdded</strong> field in a catalog snapshot dated October 30, 2026: 187 in 2023, 186 in 2024, 245 in 2025, and 246 from January to September 2026.</p><p><em>Source: aggregated data from the </em><a href="https://github.com/cisagov/kev-data"><em>official CISA KEV dataset</em></a><em>, snapshot dated October 30, 2026. A KEV </em><em>dateAdded is the date CISA added an entry to the catalog, not necessarily the date the vulnerability was published or first exploited.</em></p><p>This comparison cuts both ways. The number of published CVEs has accelerated much faster than the number of additions to the KEV catalog. But <strong>we cannot divide these two numbers and call the result the percentage of new vulnerabilities being exploited</strong>. The KEV additions may concern CVEs from years earlier; the published CVEs belong to another population and have had different amounts of time to be exploited and investigated.</p><p>To measure an exploitation rate, we would have to follow each <em>cohort</em> of CVEs published in a given period and check which later received evidence of exploitation, accounting for the delay in observing and recording attacks. KEV is useful, but it does not capture every exploited vulnerability or the number of attacks against each one.</p><p><a href="https://www.first.org/blog/20260615-vulnerability-forecast-update">FIRST’s June analysis</a> offers a more carefully bounded observation: while disclosure volumes were growing sharply, its actionable-risk overlay — based on KEV membership and an EPSS probability above 10% — remained relatively stable during the first half of 2026. That is not proof that cyberattacks stopped growing; it is a useful warning against treating all published CVEs as equally urgent.</p><p>Nor do the absolute KEV numbers justify complacency. In the same October 30 snapshot, the count for January through October 2026 was 282, compared with 214 in the equivalent 2025 period — about a 32% increase. And October supplied a real-world counterexample: <a href="https://aws.amazon.com/blogs/security/amazon-threat-intelligence-identifies-php2shell-exploitation-prior-to-public-disclosure-in-madpot/">AWS Threat Intelligence reported PHP2Shell exploitation attempts beginning October 17</a>, four days before PHP published emergency fixes on October 21.</p><p>There is a Halloween bonus, too. Some vulnerabilities added to KEV in 2026 carry identifiers from much earlier years. The monsters are not always new: some are zombies that never quite stayed buried.</p><h3>The monster that actually grew: the patch gap</h3><figure><img alt="Diagram: the patch gap. A public fix on one cliff and an installed fix on the other, with a zombie hand rising from the chasm between them, where attackers study the public fix and build an exploit before users update." src="https://cdn-images-1.medium.com/max/1024/1*g7KcKMbEIFrykgWc8d-wWA.png" /></figure><p>There is a more concrete emerging danger in this data: the time between a fix becoming visible and reaching users. This is the <strong>patch gap</strong>. The gap is not new, and it does not end merely when a release is announced; users and organizations still have to deploy the update.</p><p>The same models that find bugs can examine patches. <a href="https://red.anthropic.com/2026/n-days/">Anthropic’s Frontier Red Team reported</a> that Mythos Preview produced a working exploit for a Firefox vulnerability within an hour of its fix landing in public source code, when the corresponding release was still 18 days away.</p><p>On October 29, <a href="https://epoch.ai/publications/open-weight-models-patch-gap">Epoch AI published a direct experimental study</a> using <strong>GLM 5.3</strong>, an open-weight model. It reported successful lab attacks against six Firefox releases between September 17 and October 26 by reasoning from public fixes. The marginal attack cost was approximately <strong>$1,000 per Firefox version</strong>, at API prices, <em>after</em> the researchers had built their harness. The full research effort cost over $100,000. The attacks also relied on a separate, previously unpublished vulnerability, and the paper does not establish that the same method succeeds universally against all browsers or configurations.</p><p>Vendors are responding. <a href="https://blog.mozilla.org/en/firefox/privacy-security/defending-patch-gap/">Mozilla says</a> Firefox has accelerated shipping so that new builds reach users <strong>weekly</strong>, while also holding certain regression tests back until fixes ship, scanning new commits for bugs, and exploring ways to stagger fixes across separate sandbox boundaries. A <a href="https://blog.mozilla.org/sumo/2026/08/19/firefox-new-release-cadence-and-what-to-expect/">Mozilla Support post from August</a> described the earlier move to a two-week major-release cycle; the October security update describes a faster current build cadence. WordPress’s automation of its security-release workflow is another response to the pressure to move faster.</p><p>Here, faster discovery is a double-edged sword. It helps defenders retire vulnerabilities sooner — but it can also make partially disclosed fixes easier to weaponize before the repaired software is installed.</p><h3>Scorecard</h3><figure><img alt="Scorecard table: eight claims from this article and what the available evidence supports for each. Supported: October 2026 had concentrated releases, disclosures surged before October, AI-assisted research finds long-lived bugs faster, counting changes inflate CVE totals, and the patch gap is easier to exploit with AI. Not established: lower exploitation rate, less secure new software, users being safer overall." src="https://cdn-images-1.medium.com/max/1024/1*EGHziYuzrCeq-lTmmhp6ZQ.png" /></figure><h3>So, haunted or not?</h3><p>My reading is that the perception of a software security apocalypse is, at least partly, a measurement problem. We are exposing defects that spent years in the dark, expanding the number of organizations that report them, and forcing maintainers to rethink how they validate and ship fixes. That can be good news — provided disclosure is matched by effective remediation.</p><p>But this is not permission to relax. A fix published upstream is not the same as a fix installed on a vulnerable system. And once a public patch identifies a useful weakness, an attacker may now have tools capable of turning that information into a working exploit far faster than before.</p><p>The practical advice is unglamorous. Enable automatic updates where appropriate; for systems you manage, combine CISA KEV, EPSS, exposure, asset criticality, and vendor guidance instead of prioritizing by raw CVE count alone. Track whether updates actually reach production, not merely whether a ticket was filed.</p><p>Better flashlights are finally showing us the monsters that were already in the house. The scary part is not how many we found this October. It is how long we leave the door open after we know where they are.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1234898a6adb" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Exploring Nuxt 3  Middlewares: A Comprehensive Overview]]></title>
            <link>https://figueiredoluiz.medium.com/exploring-nuxt-3-middlewares-a-comprehensive-overview-6c45635cd6e4?source=rss-67f54c9d481d------2</link>
            <guid isPermaLink="false">https://medium.com/p/6c45635cd6e4</guid>
            <category><![CDATA[vue-3]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[nuxt-3]]></category>
            <category><![CDATA[nuxt]]></category>
            <category><![CDATA[typescript]]></category>
            <dc:creator><![CDATA[Luiz Figueiredo]]></dc:creator>
            <pubDate>Mon, 06 Mar 2023 17:32:52 GMT</pubDate>
            <atom:updated>2023-06-25T02:49:39.662Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*KVYKz-XyhECWOcd68ZVkHQ.jpeg" /></figure><h3>Exploring Nuxt 3 Middleware: A Comprehensive Overview</h3><p>Nuxt.js Middlewares are custom functions run before rendering a page or group of pages. Checking authentication, modifying the server response, or accessing cookies are some of the most common usages of middleware.</p><h3>Rendering modes</h3><p>Before talking more about middleware, let’s talk about Nuxt.js rendering modes.</p><p>Client-side Only Rendering is a mode where a Vue.js application is rendered in the browser (or client), generates HTML elements after the browser downloads, and parses all the JavaScript code.</p><p>Universal Rendering involves the Server-side and Client-side Rendering of a webpage. In Universal Rendering, the server returns a fully rendered HTML page to the browser, and the JavaScript (Vue.js) code is executed in a server environment, producing an HTML document. The client then loads the JavaScript code, and Vue.js takes control of the document and enables an interactivity called “hydration.”</p><p>Nuxt.js still has other rendering modes, like Hybrid, Rendering on CDN Edge Workers, and a very granular setting control using Route Rules. Some of these are still to be officially released.</p><h3>Types of Middleware</h3><p>There are two types of middleware in Nuxt.js: Route Middleware executed in the Vue part of the application and Normal Middleware executed in the Nitro part.</p><p>Nuxt 3 divides the route middleware into three types: Anonymous (or inline), Named Route, and Global.</p><p>The most popular use case for Nuxt middleware is the authentication guard. Still, some use cases for server middleware in Nuxt 3 include matching route parameters, accessing request cookies, logging requests, and warming up an API.</p><p>You define the Anonymous (or Inline) Route Middleware directly on the page.</p><p>An example of a page file defining middleware:</p><pre>&lt;script setup&gt;<br>definePageMeta({<br>  middleware: defineNuxtRouteMiddleware((to, from) =&gt; {<br>    console.log(&#39;to&#39;, to);<br>    console.log(&#39;from&#39;, from);<br>  }),<br>});<br>&lt;/script&gt;<br>&lt;template&gt;<br>  …<br>&lt;/template&gt;</pre><p>Named route middleware should be placed in the <em>middleware/</em> directory and will be automatically loaded via asynchronous import when used on a page.</p><p>Example of a <em>middleware/auth.ts</em>:</p><pre>export default defineNuxtRouteMiddleware(async (to, from) =&gt; {<br>  console.log(&#39;middleware/auth&#39;)<br>  console.log(&#39;to&#39;, to)<br>  console.log(&#39;from&#39;, from)<br>})</pre><p>Global route middleware is placed in the <em>middleware/</em> directory (with a <em>.global</em> suffix) and will be automatically run on every route change.</p><p>Example of a <em>middleware/auth.global.ts</em>:</p><pre>export default defineNuxtRouteMiddleware(async (to, from) =&gt; {<br>  console.log(&#39;middleware/auth.global&#39;)<br>})</pre><p>As you can see, there is no difference between their syntax. Still, they behave slightly differently, being the global executed on accessing any route, even not being directly called.</p><h3>A Named Route Middleware</h3><p>Supposing you have a file structure as proposed in the example above, your authentication guard middleware would be at <em>middleware/auth.ts</em>.</p><p>In your page file, you can reference this route middleware like this:</p><pre>&lt;script setup&gt;<br>  definePageMeta({<br>    middleware: [&quot;auth&quot;]<br>    // or middleware: &#39;auth&#39;<br>  })<br>&lt;/script&gt;</pre><p>One of the most significant advantages of this method is that you have better control over what pages would require the middleware.</p><h3>Controlling when the Middleware Runs</h3><p>If your application runs on SSR mode, Nuxt will execute the middleware for the initial page both when rendering the page and then again on the client side. Using the process object, you can control the code execution time.</p><p>The code below is from Nuxt.js official documentation and shows some possibilities for controlling the runtime.</p><pre>export default defineNuxtRouteMiddleware(to =&gt; {<br>  // skip middleware on server<br>  if (process.server) return<br>  // skip middleware on client side entirely<br>  if (process.client) return<br>  // or only skip middleware on initial client load<br>  const nuxtApp = useNuxtApp()<br>  if (process.client &amp;&amp; nuxtApp.isHydrating &amp;&amp; nuxtApp.payload.serverRendered) return<br>})</pre><h3>Parameters overview</h3><p>Route middleware is directly related to navigation. They act between the current route and the following route so that you can access this information in their arguments.</p><pre>export default defineNuxtRouteMiddleware((to, from) =&gt; {<br>  if (to.params.id === &#39;1&#39;) {<br>    return abortNavigation()<br>  }<br>  return navigateTo(&#39;/&#39;)<br>})</pre><p>Nuxt can return two globally available helpers directly from the middleware:</p><ul><li><em>navigateTo</em> can be used to redirect to a given route. It can be handy within plugins or middleware where you must programmatically change the user’s location. Additionally, you can also call <em>navigateTo</em> directly to perform page navigation.</li><li><em>abortNavigation</em>, allows you to stop navigation altogether with an optional error message. It can be helpful in cases where you need to prevent the user from navigating to a specific page or section of your application.</li></ul><h4>Read More</h4><ul><li>Middleware Directory → <a href="https://nuxt.com/docs/guide/directory-structure/middleware">https://nuxt.com/docs/guide/directory-structure/middleware</a></li><li><em>abortNavigation</em> helper → <a href="https://nuxt.com/docs/api/utils/abort-navigation">https://nuxt.com/docs/api/utils/abort-navigation</a></li><li><em>navigateTo</em> helper → <a href="https://nuxt.com/docs/api/utils/navigate-to">https://nuxt.com/docs/api/utils/navigate-to</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6c45635cd6e4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Navigating Japan’s IT Industry: A Personal Perspective]]></title>
            <link>https://figueiredoluiz.medium.com/navigating-japans-it-industry-a-personal-perspective-e9b35bfe69c?source=rss-67f54c9d481d------2</link>
            <guid isPermaLink="false">https://medium.com/p/e9b35bfe69c</guid>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[expat]]></category>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[japan]]></category>
            <category><![CDATA[japanese-culture]]></category>
            <dc:creator><![CDATA[Luiz Figueiredo]]></dc:creator>
            <pubDate>Fri, 03 Mar 2023 18:29:41 GMT</pubDate>
            <atom:updated>2023-03-03T18:29:41.586Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5tzHxmKk7XY8DHvC_NHSvA.jpeg" /><figcaption>Osaka, Dotonbori</figcaption></figure><p>Living in Japan for an extended period has given me a unique insight into the country’s IT industry. Initially, moving to Japan was a challenging decision, but I was determined to make it work. Even though I had studied Japanese before moving, I realized that I needed more to establish proper communication with the locals. I had to take a moment of dense study and force myself to practice as much as possible, even taking jobs out of my career path to get more practice.</p><p>I was unprepared, and it took me about six months to get a job as a developer. Despite some companies being foreign-friendly and offering all the support for settling in Japan, you still need to face processes, documents, and bureaucratic barriers. The salaries in Japan are not the highest in the world, absolutely less than in the US, and not always more than in South America. The trade-off is the country’s infrastructure and security, which makes a big difference in your everyday life.</p><p>There is still a need for more Engineers, especially Programmers, in Japan. The more you can speak Japanese, the more opportunities you get. However, finding a company willing to hire non-Japanese speakers can be challenging. You may have to search for foreign-friendly companies or improve your language skills.</p><p>The average IT companies in Japan are not innovation-oriented, as we assume, and many areas resist adopting new technologies. Most companies are resistant to renovating the code. If your field is related to web development, you can face 10 to 15-year-old source codes still being maintained. In my experience, I developed a web chat and an online notification system for a horse bet website that still had code from the early 2000s.</p><p>Recently, I have noticed some changes in the industry in Japan, and better startups have been born. Many companies are starting to adopt new programming languages and architectures. Some modern work formats are also becoming more common, and the resistance to remote work has reduced significantly. It is clear that the industry is evolving, and it’s an exciting time to be a part of it.</p><p>Suppose you want to know more about the country or consider it a place to work. In that case, you should check different media sources and visit the country before deciding. It’s important to note that there are many myths about Japan, and you should never take one person’s point of view as the absolute truth.</p><p>In conclusion, Japan’s IT industry has unique challenges. Japan’s infrastructure and security make it an attractive place to live. Although the salaries may not be the highest, the trade-off is worth considering. Still, it’s possible to succeed with hard work, dedication, and a willingness to learn. Knowing the language, adaptability, and embracing the culture are keys to making it work.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e9b35bfe69c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>