<?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 Gifa Eriyanto on Medium]]></title>
        <description><![CDATA[Stories by Gifa Eriyanto on Medium]]></description>
        <link>https://medium.com/@gifaeriyanto?source=rss-38918a87a0f5------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*JGMkt9-V7y6zzRNtNgDe-A@2x.jpeg</url>
            <title>Stories by Gifa Eriyanto on Medium</title>
            <link>https://medium.com/@gifaeriyanto?source=rss-38918a87a0f5------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 07:50:07 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@gifaeriyanto/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[One Engineer’s 287,748 Agent Calls Suggest Compacting at 250K–450K Tokens]]></title>
            <link>https://gifaeriyanto.medium.com/one-engineers-287-748-agent-calls-suggest-compacting-at-250k-450k-tokens-d72161dd1039?source=rss-38918a87a0f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/d72161dd1039</guid>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[context-engineering]]></category>
            <dc:creator><![CDATA[Gifa Eriyanto]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 13:08:44 GMT</pubDate>
            <atom:updated>2026-11-15T13:08:44.017Z</atom:updated>
            <content:encoded><![CDATA[<h3>A vendor’s case study on when to compact a coding agent, read with its limits in plain view.</h3><figure><img alt="A small brass robot sits beside a towering stack of paper scrolls and an analog meter in a warm attic workshop overlooking a green valley." src="https://cdn-images-1.medium.com/max/1024/1*QKMYfBi0XU7qXrayCiMTTw.png" /><figcaption>Original illustration: as context piles up, the meter keeps running.</figcaption></figure><p>Your coding agent has a habit that is easy to forget: on every step, it re-sends everything it has seen so far.</p><p>The first step carries a little context. The fiftieth carries the whole conversation plus everything that happened along the way. Nobody sees that on the screen, but it is what the meter is counting. So the question for anyone running long sessions is a plain one. What does the window cost as it grows, and what do you lose when you compact it back down?</p><p>One research report tries to answer that with real data. It is worth reading carefully, partly because the numbers are striking and partly because the author is upfront about where they stop applying. This piece is about both halves.</p><p><em>(This is a follow-up to </em><a href="https://gifaeriyanto.medium.com/why-coding-agents-burn-money-what-a-50-turn-session-could-cost-34ef0e946585"><em>Why Coding Agents Burn Money: What a 50-Turn Session Could Cost</em></a><em>, and to Sunday’s piece on cost per task. You don’t need either to follow this one.)</em></p><h3>Who ran this, and what it is</h3><p>The report is “Finding the Optimal Context Window for Coding Agents: A Case Study” by Rogerio Chaves at LangWatch, published August 2, 2026. LangWatch sells LLM observability, which is a reason to read it as a vendor’s research and not a neutral lab’s. It isn’t peer reviewed.</p><p>The dataset, in the report’s own summary line: “287,748 API calls · 2,451 agents · 162 days · 873 compactions · 201 audited steps.”</p><p>Here is the caveat that matters most, and the report says it itself: “Single subject. This is one engineer’s usage, weighted heavily toward agentic coding on a specific harness and model family. The constants are theirs.” The blog version repeats it: “All of this is my own usage: mostly agentic coding, mostly on the LangWatch codebase. Your constants will differ.”</p><p>The report doesn’t name the harness or the model family it measured, so I won’t either. The only model it names is the one that judged the audited steps, an Opus judge. And everything is measured in tokens, not dollars. Keep that in mind for the rest of the article.</p><h3>Finding one: the cost of getting big</h3><p>The report measures what it takes to <em>reach</em> a given context size, meaning the tokens consumed along the way across all the steps, not the price of one call at that size. Its numbers: “Reaching 100k of context costs a median of 1.4M tokens. Reaching 200k costs 9M. Reaching 800k costs 338M, which is 250 times the 100k figure for eight times the window.”</p><p>The fit has an exponent of 2.55 (R² = 0.994), which the report summarizes as “each doubling of window size costs roughly six times more.” Notice “roughly.” The exponent was measured on this one subject, and a refit moves it somewhat, so I’d treat six as a rule of thumb and not a law.</p><figure><img alt="A log-scale bar chart shows 1.4M cumulative tokens to reach 100K context, 9M tokens to reach 200K context, and 338M tokens to reach 800K context." src="https://cdn-images-1.medium.com/max/1024/1*PQb8YD6mY9TEx0VRnL8Vwg.png" /><figcaption>Source: LangWatch Research, Aug 2, 2026. One engineer&#39;s usage; tokens, not dollars. Median cumulative tokens to reach each context size.</figcaption></figure><p>What about caching? That’s the obvious objection: cached reads are cheaper. The report agrees and then adds the catch. On the incident day it highlights, “cache reads were 95% of all tokens consumed. A cached token is still a token purchased on every subsequent step. Caching multiplies the constant and leaves the exponent untouched.” The blog adds that cache reads are “about 10x cheaper per token,” but still made up that 95%. In other words, caching changes how steep the bill is at any given size. It doesn’t change the shape of the curve.</p><h3>A small illustration, labeled as one</h3><p>To see why “cheap per token” and “big in total” can both hold, here is a made-up session. <strong>This is illustrative arithmetic from published rates. It is not from the study and not a real session cost.</strong></p><p>Assumptions: Claude Fable 5.2, a 200,000-token context re-sent over 50 steps, every input read coming from cache at $0.20 per million tokens, cache-write costs and output ignored.</p><ul><li>Tokens read: 200,000 × 50 = 10,000,000, which at $0.20 per million is <strong>$2.00</strong>.</li><li>If half of that context were <em>not</em> cached and billed at the standard $8 per million input rate: 5,000,000 tokens × $8 per million = $40.00, plus the other 5,000,000 cached at $0.20 per million = $1.00, for <strong>$41.00</strong>.</li></ul><p>The point isn’t the dollar figure. It’s that a re-read can be cheap per token and still be most of what you consume. Please don’t mix these numbers with the study’s, which count tokens only.</p><figure><img alt="A two-bar chart compares operator corrections at 17.7% for the within-session baseline and 41.9% during the five steps after compaction." src="https://cdn-images-1.medium.com/max/1024/1*H1loW5k4jtoJQoxp0OL5Eg.png" /><figcaption>Source: LangWatch Research, Aug 2, 2026. One engineer&#39;s usage. Observational, no significance test reported.</figcaption></figure><h3>Finding two: what compacting takes away</h3><p>If big windows are expensive, the instinct is to compact early. The report’s second finding is the reason it isn’t free.</p><p>Compaction replaces your history with a summary. In the report’s words: “the median event reduces a 574,875-token window to a 4,382-token summary. Slightly under 1% of the raw material survives. Each event costs a median of 142 seconds of wall clock.”</p><p>And then the part I found most useful. The share of operator messages that contain corrections rises to 41.9% in the five steps after a compaction, against a 17.7% baseline within the same sessions. The blog calls that 2.37 times, says it “stays elevated for at least 30 steps,” and that it “has not fully recovered by 120.”</p><p>Read that with some care. The correction measure is a rough proxy, built from phrase matching. The report also admits a reverse-causation problem: compaction is partly discretionary, so an operator may compact <em>because</em> a session is already going badly, which would raise the correction rate without compaction doing anything. The report says that a rise lasting beyond 100 steps is hard to explain that way, but it states the measurement is observational and doesn’t claim cause. It also reports no significance tests or confidence intervals for this ratio. The only fit statistic it gives is the R² on the cost curve.</p><p>So the careful reading is: after a compaction, this engineer’s sessions looked worse for a long while, and we can’t yet say compaction was the cause.</p><p>One more number is worth holding on to. The auditors found that “28 of 195 steps, 14%, genuinely required verbatim old context.” Most of the time, the agent didn’t need the old text word for word. When it did, a summary couldn’t give it back.</p><h3>So when should you compact?</h3><p>This is where it’s easy to quote the wrong number, so let me be exact.</p><p>The report’s abstract gives “an operating range of 250,000 to 450,000 tokens differentiated by task type rather than a single threshold.” But that isn’t where its first analysis lands. The report says that “taken in isolation,” the cost and decay sections “converge on a single recommendation: compact early, in the region of 150k to 220k.” The higher range comes after the report weighs four considerations together, including the behavior cost after compaction. So the 150K–220K range is the cost view on its own, and 250K–450K is the combined view. Neither one is a universal threshold.</p><p>The report then splits by task type. These are its thresholds, with its own short rationale:</p><figure><img alt="A table shows suggested compaction ranges of 200K–250K for driving a PR, 250K–300K for research, 250K–350K for QA and dogfooding, and 300K–450K for both building and code understanding, with abridged report notes." src="https://cdn-images-1.medium.com/max/1024/1*eT1NPxChDNIFOOPrf08hEw.png" /><figcaption>Source: LangWatch Research, Aug 2, 2026, Table 2 (notes abridged). One engineer&#39;s usage; constants will differ.</figcaption></figure><p>Two things I’d take from this table, neither of which is a number. Work that depends on exact strings, such as building, tolerates a bigger window because a summary can’t hold the exact text. Self-contained work like research can be compacted sooner because there’s less to lose. That logic should transfer further than the specific thresholds do.</p><h3>Is anyone else saying this?</h3><p>Not the numbers, no. I found no independent replication of the LangWatch figures.</p><p>What exists is support for the general direction. Chroma’s July 2025 research “Context Rot” evaluated 18 models and found that “models do not use their context uniformly; instead, their performance grows increasingly unreliable as input length grows.” It has limits of its own: Chroma sells a vector database, the tasks weren’t coding agents, and the models are from 2025. Anthropic’s own documentation describes compaction as keeping “the active context small, because response quality degrades as a conversation grows.” That confirms why the feature exists. It gives no threshold.</p><p>So the direction is plausible and consistent across sources. The exact range is one person’s.</p><h3>How you could test it yourself</h3><p>I haven’t tried any of this, and I haven’t compacted at 250K on my own sessions, so this is a suggestion and not a result.</p><p>If you run long sessions, you could log two things for a few weeks. First, the token count at each point you compact or restart. Second, how often you have to correct the agent in the five steps afterward, compared with your normal rate. If your own numbers look like the report’s, you’ve learned something about your work. If they don’t, you’ve learned your constants differ, which is exactly what the author predicted.</p><p>Whatever you find, keep it separate from the dollar bill. The study counts tokens, and so has everything in this piece.</p><p><em>This article was researched and drafted with AI assistance, then reviewed and edited by a human before publishing.</em></p><p><em>Sources: </em><a href="https://langwatch.ai/research/finding-the-optimal-context-window"><em>LangWatch research report</em></a><em> · </em><a href="https://langwatch.ai/blog/context-tax-when-to-compact"><em>LangWatch blog post</em></a><em> · </em><a href="https://research.trychroma.com/context-rot"><em>Chroma, Context Rot</em></a><em> · </em><a href="https://platform.claude.com/docs/en/build-with-claude/compaction"><em>Anthropic compaction docs</em></a><em> · </em><a href="https://platform.claude.com/docs/en/about-claude/pricing"><em>Claude pricing docs</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d72161dd1039" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Coding Agents Burn Money: What a 50-Turn Session Could Cost]]></title>
            <link>https://gifaeriyanto.medium.com/why-coding-agents-burn-money-what-a-50-turn-session-could-cost-34ef0e946585?source=rss-38918a87a0f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/34ef0e946585</guid>
            <category><![CDATA[cost-optimization]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[programming]]></category>
            <dc:creator><![CDATA[Gifa Eriyanto]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 09:37:16 GMT</pubDate>
            <atom:updated>2026-11-14T09:37:16.566Z</atom:updated>
            <content:encoded><![CDATA[<p>Coding agent cost often hides in failures and retries. The TokenEconomizer paper (Microsoft Research and UC Berkeley, arXiv 2610.22732, https://arxiv.org/abs/2610.22732) profiles 60 SWE-Bench tasks, which are real GitHub issue fixes, 10 from each of six repositories. Within that sample, the paper reports: &quot;The 10 failed executions constitute 17% of the tasks but 32% of total spend (Fig. 3(c)).&quot; This is a retrospective profile of one sample, not a general statistic. Still, the pattern is worth noticing, because failed runs are easy to overlook while the answers scroll by.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TNbVZNcvQbS89AwqaD72Ow.png" /><figcaption>Illustration for this article.</figcaption></figure><blockquote>TL;DR: Coding agent cost is driven mostly by rereading accumulated context on every turn, through cache reads (reusing stored context) and cache writes (storing it), not by the output the agent writes. One paper measures these shares directly. In PointFive&#39;s analysis of 2,848 Claude Code runs, cache writes made up 48.5% of cost, cache reads 38.7% and output 11.4%. Another paper finds that input tokens dominate. A 50-turn model below shows a similar pattern. It is illustrative, not measured.</blockquote><h3>Where does a coding agent&#39;s money go?</h3><p>A coding agent resends its history on every turn. That history is usually cached, meaning the provider stores it so later requests can reuse it at a lower price. A request served from the cache is a &quot;cache hit.&quot; A request the cache cannot serve is a &quot;cache miss,&quot; and it is billed at full input. Rereads are cheap per token, but they add up fast. The pricing pages show the structure.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qJRit4fPllNoR3sk2q56Cg.png" /><figcaption>Claude API pricing in US dollars per million tokens, as shown on the Anthropic pricing page on November 14, 2026. Source: https://platform.claude.com/docs/en/about-claude/pricing (accessed November 14, 2026).</figcaption></figure><p>For Opus 5.5, the pricing page lists base input at $4 per million tokens and output at $20. A cache read costs $0.20, which is 5% of base input. Writing to a 5-minute cache costs $5, or 1.25x input. A 1-hour cache write costs $8, or 2x. Caching is cheap to read and slightly more expensive to write, so it pays off only when the same context is read many times.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lfuWoGBWGy_kPgLvgf9UHQ.png" /><figcaption>Prompt caching multipliers on the Anthropic pricing page, accessed November 14, 2026. Source: https://platform.claude.com/docs/en/about-claude/pricing.</figcaption></figure><p>DeepSeek is priced far lower. Its Flash model lists a peak-hour cache-hit price of $0.006 per million tokens, a cache miss at $0.30 and output at $1.20. Off-peak prices are half those rates. V4 Pro costs more: $0.044 for a cache hit, $1.32 for a cache miss and $3.96 for output at peak.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BCxdFBXSpMDi4Rv4NiJx5Q.png" /><figcaption>DeepSeek API pricing, accessed November 14, 2026. Source: https://api-docs.deepseek.com/quick_start/pricing. DeepSeek reserves the right to change its prices.</figcaption></figure><h3><strong>Does lower output mean a lower bill?</strong></h3><p>The clearest breakdown comes from Weinberger and Hozez at PointFive, &quot;Token Reduction Is Not Cost Reduction&quot; (arXiv 2607.12161, https://arxiv.org/abs/2607.12161). The paper analyzed 2,848 Claude Code runs and split the bill this way: uncached input 1.4%, cache writes 48.5%, cache reads 38.7% and output 11.4%. Output is a small share. Most of the money goes to writing and rereading context.</p><p>The same paper shows why cutting tokens is not the same as cutting cost. The most aggressive compression of tool output cut tool-output tokens by 38.4%, yet raised the bill by 6.8%. Across runs, the correlation between token reduction and cost reduction was only r = 0.15. The study covers a single harness and a single provider, mostly on single-prompt tasks, and the authors say so.</p><h3>How much do repeated runs vary?</h3><p>Variance is the other problem. &quot;How Do AI Agents Spend Your Money?&quot; (arXiv 2604.22750, https://arxiv.org/abs/2604.22750) studied OpenHands on SWE-bench Verified across eight models. Its abstract puts agentic tasks at roughly 1,000 times the tokens of code reasoning or chat. The body reports larger multiples, 3,500 and 1,200, so the exact figure depends on the comparison. According to the paper, repeated runs of the same task can differ by up to 30 times in total tokens. It also finds that input tokens, not output, dominate cost even when caching is enabled. The models it studied are older, and the authors note their limits.</p><h3>Does the 31% saving hold up?</h3><p>The headline &quot;31% cheaper&quot; comes from TokenEconomizer (arXiv 2610.22732). Its exact terms matter.</p><p>Across 277 SWE-Bench tasks, the plugin cut average cost per task by between 18.4% and 30.9%, counting the plugin&#39;s own call costs. Across all 579 tasks in the paper, which include Web-Bench and security tasks, the reported ranges are wider. Pass@1 (the share of tasks solved on the first attempt) rose in four of five agent-model combinations. The 30.9% figure is the top of the SWE-Bench range, and it applies to one combination only: Claude Code with Opus 4.8. The paper did not test Opus 5.5 or Sonnet 5.5, so &quot;up to 30.9%&quot; is not a promise about current Claude models.</p><p>The analysis suggests trusting the direction, not the number. The PointFive study&#39;s tool-output finding shows that a token cut does not guarantee a bill cut. Before relying on any percentage, run the same tasks on your own bill.</p><h3>What could a 50-turn session cost?</h3><p>Even with a cache, a hypothetical 50-turn session reads about 5.5 million tokens from cache, writes about 200,000 and produces about 20,000 output tokens. The model below shows the structure of cost in that case. It is not a measurement of any real session. Prices were accessed November 14, 2026.</p><p>The model rests on these assumptions:</p><ul><li>Starting context of 25,000 tokens: system prompt, tool schemas and repository notes.</li><li>Each turn adds 3,500 tokens of tool results and replies.</li><li>Each turn produces 400 output tokens.</li><li>50 turns in total, with no failures, no subagents and no context above 200K tokens.</li><li>Every turn reads the full history from cache. The model assumes the cache never expires and never misses. Real cache behavior is not confirmed here. The cache&#39;s time-to-live (TTL, how long an entry stays valid before it expires) and any invalidation would both affect the result, and misses would raise the cost.</li></ul><p>Illustrative model, not measured data. Prices accessed November 14, 2026.</p><ul><li>Fable 5.2: $3.90 with cache, $46.50 without cache. Share of cached cost: 28% reads / 51% writes / 20% output.</li><li>Opus 5.5: $2.50 with cache, $23.25 without cache. Share of cached cost: 44% reads / 40% writes / 16% output.</li><li>Sonnet 5.5: $1.25 with cache, $11.63 without cache. Share of cached cost: 44% reads / 40% writes / 16% output.</li></ul><p>Under the same assumptions, DeepSeek Flash costs about $0.12 at peak and $0.06 off-peak. V4 Pro costs about $0.59 at peak and $0.29 off-peak. These figures show shape, not what a real session will cost.</p><p>In this illustrative model, the same session without the cache costs about 9 to 12 times more on the Claude models. Output is also a small share of the total. A shorter session shifts the mix: in a lighter 30-turn scenario, output reaches roughly a third of the total. In this model, session length matters more than reply length.</p><h3>What actually lowers the bill?</h3><p>The levers below come from the pricing pages and the research. The papers have not measured how much each one saves in a real session.</p><ol><li><strong>Use the cheaper model for light steps.</strong> The Sonnet 5.5 price is half the Opus 5.5 price in every column of the table: input $2 against $4, output $10 against $20, cache read $0.10 against $0.20.</li><li><strong>Use batch processing for work that can wait.</strong> Batch (discounted, asynchronous processing) for Opus 5.5 is $2 input and $10 output, half the standard rate.</li><li><strong>Skip fast mode unless you need it.</strong> Fast mode doubles the Opus 5.5 price to $8 input and $40 output.</li><li><strong>Keep sessions short, or start fresh on purpose.</strong> Every extra turn adds to the context that gets reread. A new session discards that history, which can help the bill more than trimming replies.</li><li><strong>Test cheaper models on your own tasks.</strong> DeepSeek Flash&#39;s peak cache-miss price of $0.30 per million tokens is far below Claude&#39;s. The price alone does not show equal quality, so test before switching.</li></ol><p>No measured saving exists for any of these levers. Treat them as direction for your own bill.</p><h3>Is there a fair way to measure coding agent cost?</h3><p>No public standard yet measures the number that matters most: cost per finished task, not cost per token. Until one exists, the most useful habit is to compare your own bill, per completed task, before and after each change.</p><p><em>This article was researched and drafted with AI assistance, then reviewed and edited by a human before publishing.</em></p><h3>Sources (accessed November 14, 2026)</h3><ul><li>Anthropic API pricing: <a href="https://platform.claude.com/docs/en/about-claude/pricing">https://platform.claude.com/docs/en/about-claude/pricing</a></li><li>DeepSeek API pricing: <a href="https://api-docs.deepseek.com/quick_start/pricing">https://api-docs.deepseek.com/quick_start/pricing</a></li><li>TokenEconomizer, arXiv 2610.22732: <a href="https://arxiv.org/abs/2610.22732">https://arxiv.org/abs/2610.22732</a></li><li>How Do AI Agents Spend Your Money?, arXiv 2604.22750: <a href="https://arxiv.org/abs/2604.22750">https://arxiv.org/abs/2604.22750</a></li><li>Token Reduction Is Not Cost Reduction, arXiv 2607.12161: <a href="https://arxiv.org/abs/2607.12161">https://arxiv.org/abs/2607.12161</a></li></ul><p>Prices were accessed on November 14, 2026. DeepSeek reserves the right to change its prices.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=34ef0e946585" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Stop Correcting the Same Claude Code Session. Hand It Off Instead.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://gifaeriyanto.medium.com/stop-correcting-the-same-claude-code-session-hand-it-off-instead-59f38880e5f9?source=rss-38918a87a0f5------2"><img src="https://cdn-images-1.medium.com/max/1536/1*EwNg3o5j70FCcXLJaB-Itw@2x.jpeg" width="1536"></a></p><p class="medium-feed-snippet">You&#x2019;re probably wasting hours repeating the same instructions.</p><p class="medium-feed-link"><a href="https://gifaeriyanto.medium.com/stop-correcting-the-same-claude-code-session-hand-it-off-instead-59f38880e5f9?source=rss-38918a87a0f5------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://gifaeriyanto.medium.com/stop-correcting-the-same-claude-code-session-hand-it-off-instead-59f38880e5f9?source=rss-38918a87a0f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/59f38880e5f9</guid>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Gifa Eriyanto]]></dc:creator>
            <pubDate>Fri, 07 Aug 2026 21:47:08 GMT</pubDate>
            <atom:updated>2026-08-07T21:47:08.075Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Matt Pocock’s /teach Skill Changed How I Learn With AI Agents]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://gifaeriyanto.medium.com/matt-pococks-teach-skill-changed-how-i-learn-with-ai-agents-cee9df5f807f?source=rss-38918a87a0f5------2"><img src="https://cdn-images-1.medium.com/max/1536/1*nadUpBBRPwCrg_DNtv3xmg.png" width="1536"></a></p><p class="medium-feed-snippet">I did not expect to become a fan of an AI skill.</p><p class="medium-feed-link"><a href="https://gifaeriyanto.medium.com/matt-pococks-teach-skill-changed-how-i-learn-with-ai-agents-cee9df5f807f?source=rss-38918a87a0f5------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://gifaeriyanto.medium.com/matt-pococks-teach-skill-changed-how-i-learn-with-ai-agents-cee9df5f807f?source=rss-38918a87a0f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/cee9df5f807f</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[learning]]></category>
            <dc:creator><![CDATA[Gifa Eriyanto]]></dc:creator>
            <pubDate>Wed, 15 Jul 2026 12:01:01 GMT</pubDate>
            <atom:updated>2026-07-15T12:01:01.531Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[OpenAI Temporarily Removed Codex’s Five-Hour Limit. Is That Good News?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://gifaeriyanto.medium.com/openai-temporarily-removed-codexs-five-hour-limit-is-that-good-news-c504786b2be6?source=rss-38918a87a0f5------2"><img src="https://cdn-images-1.medium.com/max/1672/1*v-iMuuGAKOnFG4-Slx72bQ.png" width="1672"></a></p><p class="medium-feed-snippet">OpenAI paused the cap for Plus, Business, and Pro, but weekly limits&#x2014;and GPT-5.6 Sol usage issues&#x2014;remain.</p><p class="medium-feed-link"><a href="https://gifaeriyanto.medium.com/openai-temporarily-removed-codexs-five-hour-limit-is-that-good-news-c504786b2be6?source=rss-38918a87a0f5------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://gifaeriyanto.medium.com/openai-temporarily-removed-codexs-five-hour-limit-is-that-good-news-c504786b2be6?source=rss-38918a87a0f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/c504786b2be6</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Gifa Eriyanto]]></dc:creator>
            <pubDate>Mon, 13 Jul 2026 21:35:27 GMT</pubDate>
            <atom:updated>2026-07-13T21:35:27.181Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[AI Coding Mistakes #2: Copy-Pasting AI Code Without Understanding It]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://levelup.gitconnected.com/ai-coding-mistakes-2-copy-pasting-ai-code-without-understanding-it-d3204517100e?source=rss-38918a87a0f5------2"><img src="https://cdn-images-1.medium.com/max/1536/1*ioHHH4S8JoBiIrDEPKCoHw@2x.jpeg" width="1536"></a></p><p class="medium-feed-snippet">Most developers don&#x2019;t realize they stopped understanding their own code</p><p class="medium-feed-link"><a href="https://levelup.gitconnected.com/ai-coding-mistakes-2-copy-pasting-ai-code-without-understanding-it-d3204517100e?source=rss-38918a87a0f5------2">Continue reading on Level Up Coding »</a></p></div>]]></description>
            <link>https://levelup.gitconnected.com/ai-coding-mistakes-2-copy-pasting-ai-code-without-understanding-it-d3204517100e?source=rss-38918a87a0f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/d3204517100e</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[career-advice]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Gifa Eriyanto]]></dc:creator>
            <pubDate>Mon, 23 Mar 2026 20:37:36 GMT</pubDate>
            <atom:updated>2026-03-23T20:37:36.259Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[AI Coding Mistakes #1: Treating AI Like Autocomplete]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://levelup.gitconnected.com/ai-coding-mistakes-1-treating-ai-like-autocomplete-6311db09ccca?source=rss-38918a87a0f5------2"><img src="https://cdn-images-1.medium.com/max/2600/0*VPP4hRABuG7YjB0B" width="6000"></a></p><p class="medium-feed-snippet">Most developers treat AI coding tools like smarter autocomplete. That habit quietly creates messy architecture and technical debt.</p><p class="medium-feed-link"><a href="https://levelup.gitconnected.com/ai-coding-mistakes-1-treating-ai-like-autocomplete-6311db09ccca?source=rss-38918a87a0f5------2">Continue reading on Level Up Coding »</a></p></div>]]></description>
            <link>https://levelup.gitconnected.com/ai-coding-mistakes-1-treating-ai-like-autocomplete-6311db09ccca?source=rss-38918a87a0f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/6311db09ccca</guid>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Gifa Eriyanto]]></dc:creator>
            <pubDate>Thu, 12 Mar 2026 23:39:40 GMT</pubDate>
            <atom:updated>2026-03-13T14:11:33.241Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Is Programming Becoming a Pay-to-Win Game?]]></title>
            <link>https://levelup.gitconnected.com/is-programming-becoming-a-pay-to-win-game-466345a1f550?source=rss-38918a87a0f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/466345a1f550</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Gifa Eriyanto]]></dc:creator>
            <pubDate>Wed, 11 Mar 2026 19:18:25 GMT</pubDate>
            <atom:updated>2026-03-13T14:17:01.535Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QwWDHTrhdpr1B430LP3x4g@2x.jpeg" /><figcaption>Photo by <a href="https://www.linkedin.com/in/zulfugarkarimov">Zulfugar Karimov</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>A few years ago, programming was mostly about skill.</p><p>Your knowledge of algorithms. Your ability to debug problems. Your experience designing systems.</p><p>But in 2025, something subtle has changed. Many developers quietly ask themselves a strange question:</p><p><em>“Am I bad at coding… or am I just using the wrong AI model?”</em></p><p>Because today, the developer with the better AI model and bigger token budget can often produce better code. And that raises an uncomfortable question:</p><p><em>Is programming slowly becoming a </em><strong><em>pay-to-win</em></strong><em> game?</em></p><h3>The New Stack: Code + AI + Budget</h3><p>Modern development increasingly relies on AI tools like Cursor, Claude Code, GitHub Copilot, and ChatGPT. These tools can generate large parts of your codebase and dramatically speed up development.</p><p>But there’s an important detail most people overlook.</p><p>The quality of AI-generated code depends heavily on three things: <strong>model quality, context size, and token budget</strong>. And all three are tied to money.</p><p>For example, a developer using a smaller model with limited context may get mediocre suggestions. Meanwhile, another developer using a more powerful model with larger context windows may receive significantly better results.</p><p>The difference is not just cleaner code. Sometimes the AI even suggests <strong>better architectural decisions</strong>.</p><h3>Context Is the Real Power</h3><p>Many developers think better AI results come from better prompts.</p><p>But the real advantage is context.</p><p>When an AI model can read your entire repository – including architecture files, documentation, modules, and previous changes – it becomes dramatically more useful. Instead of acting like autocomplete, it behaves more like a junior engineer who understands your project.</p><p>The problem is that large context windows are expensive.</p><p>Processing hundreds of thousands of tokens costs real money. Developers with larger token budgets can feed their entire codebase to the AI, while others must constantly trim context to stay within limits.</p><p>As a result, some developers are effectively working <strong>with smarter AI assistants simply because they can afford larger context usage</strong>.</p><h3>The Silent Productivity Gap</h3><p>This creates a new kind of productivity gap between developers.</p><p>Not junior vs senior. Not frontend vs backend.</p><p>Instead, it’s <strong>AI-augmented developers vs constrained developers</strong>.</p><p>Developers with powerful models, large token budgets, and agent workflows can generate and iterate on code incredibly fast. They prototype faster, refactor faster, and explore more ideas.</p><p>Meanwhile, developers using free tiers or smaller models often experience the opposite. They deal with more hallucinations, worse suggestions, and constant limitations on context size.</p><p>Two developers with similar skills may produce very different results – simply because one has <strong>better AI infrastructure</strong>.</p><h3>We’ve Seen Something Like This Before</h3><p>In some ways, technology has always had small pay-to-win elements.</p><p>Developers with faster computers compile faster. Teams with better cloud infrastructure deploy faster. Companies with better tooling often ship faster.</p><p>But those advantages were mostly about <strong>speed</strong>, not <strong>thinking</strong>.</p><p>AI tools are different. They don’t just make you faster – they amplify how you reason about code.</p><p>A strong AI coding assistant can suggest architecture patterns, generate tests, refactor modules, and identify edge cases you might miss.</p><p>That’s not just productivity.</p><p>That’s <strong>cognitive leverage</strong>.</p><h3>But Skill Still Matters</h3><p>Before we panic, there’s an important reality.</p><p>Even the most advanced AI models still struggle with ambiguous requirements, complex system design trade-offs, and long-term maintainability decisions.</p><p>AI can generate code, but it doesn’t truly understand the business context behind it.</p><p>A developer who understands architecture, debugging, testing, and system design will still outperform someone who blindly accepts AI output.</p><p>Think of AI less like autopilot and more like <strong>power steering</strong>.</p><p>It makes you stronger and faster – but it still depends on the driver.</p><h3>AI Infrastructure Is Now Part of the Developer Stack</h3><p>The real shift isn’t that programming is becoming pay-to-win.</p><p>The real shift is that <strong>AI infrastructure is becoming part of the developer stack</strong>.</p><p>Just like developers invest in better hardware, faster CI/CD pipelines, and better IDEs, many are now investing in stronger AI tooling.</p><p>That includes better models, larger context workflows, higher token budgets, and smarter integrations.</p><p>This isn’t cheating. It’s simply the next evolution of developer tooling.</p><h3>The Developers Who Will Win</h3><p>The developers who succeed in this new era won’t necessarily be the ones spending the most money.</p><p>They will be the ones who understand <strong>how to combine engineering skill with AI leverage</strong>.</p><p>They will learn how context affects AI performance. They will structure their codebase in ways that AI tools can understand more easily. And they will know when to trust AI suggestions – and when to override them.</p><p>In other words, the future isn’t pay-to-win.</p><p>It’s <strong>leverage-to-win</strong>.</p><h3>Final Thought</h3><p>Programming isn’t becoming a pay-to-win game.</p><p>But it <strong>is</strong> <strong>becoming an AI-assisted game</strong>.</p><p>And just like every tool in engineering, the real question isn’t whether you have access to AI.</p><p>The real question is:</p><p><strong><em>Do you know how to use it better than everyone else?</em></strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=466345a1f550" width="1" height="1" alt=""><hr><p><a href="https://levelup.gitconnected.com/is-programming-becoming-a-pay-to-win-game-466345a1f550">Is Programming Becoming a Pay-to-Win Game?</a> was originally published in <a href="https://levelup.gitconnected.com/">Level Up Coding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your AI Coding Tool Isn’t the Problem. Your Context Is.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://levelup.gitconnected.com/your-ai-coding-tool-isnt-the-problem-your-context-is-9659dc2005a6?source=rss-38918a87a0f5------2"><img src="https://cdn-images-1.medium.com/max/1376/1*cp1K1Cqs2cGv2Lq9zprMmw@2x.jpeg" width="1376"></a></p><p class="medium-feed-snippet">Most AI coding mistakes start long before the model writes a single line of code.</p><p class="medium-feed-link"><a href="https://levelup.gitconnected.com/your-ai-coding-tool-isnt-the-problem-your-context-is-9659dc2005a6?source=rss-38918a87a0f5------2">Continue reading on Level Up Coding »</a></p></div>]]></description>
            <link>https://levelup.gitconnected.com/your-ai-coding-tool-isnt-the-problem-your-context-is-9659dc2005a6?source=rss-38918a87a0f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/9659dc2005a6</guid>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Gifa Eriyanto]]></dc:creator>
            <pubDate>Mon, 09 Mar 2026 14:50:45 GMT</pubDate>
            <atom:updated>2026-03-09T14:50:45.002Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Code Review Like Mob Psycho 100: Power Without Ego]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://levelup.gitconnected.com/code-review-like-mob-psycho-100-power-without-ego-fefb50035437?source=rss-38918a87a0f5------2"><img src="https://cdn-images-1.medium.com/max/2600/1*f60omVsl7CaSIXaTGVBQ-g@2x.jpeg" width="2816"></a></p><p class="medium-feed-snippet">The pull request had three files changed. The junior developer waited nervously for feedback.</p><p class="medium-feed-link"><a href="https://levelup.gitconnected.com/code-review-like-mob-psycho-100-power-without-ego-fefb50035437?source=rss-38918a87a0f5------2">Continue reading on Level Up Coding »</a></p></div>]]></description>
            <link>https://levelup.gitconnected.com/code-review-like-mob-psycho-100-power-without-ego-fefb50035437?source=rss-38918a87a0f5------2</link>
            <guid isPermaLink="false">https://medium.com/p/fefb50035437</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[programming]]></category>
            <dc:creator><![CDATA[Gifa Eriyanto]]></dc:creator>
            <pubDate>Tue, 02 Dec 2025 07:55:53 GMT</pubDate>
            <atom:updated>2025-12-03T03:18:26.768Z</atom:updated>
        </item>
    </channel>
</rss>