<?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 Oleg Verhovsky on Medium]]></title>
        <description><![CDATA[Stories by Oleg Verhovsky on Medium]]></description>
        <link>https://medium.com/@oleg_verchol?source=rss-45476f21bcfc------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*66xGFpjZGAaAHEKIfjtqhg.jpeg</url>
            <title>Stories by Oleg Verhovsky on Medium</title>
            <link>https://medium.com/@oleg_verchol?source=rss-45476f21bcfc------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 03:36:11 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@oleg_verchol/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[The Agentic SDLC Iceberg: Hundreds of Agents Are Coming. Who Will Operate Them?]]></title>
            <link>https://medium.com/@oleg_verchol/the-agentic-sdlc-iceberg-hundreds-of-agents-are-coming-who-will-operate-them-4fdbfc344b2e?source=rss-45476f21bcfc------2</link>
            <guid isPermaLink="false">https://medium.com/p/4fdbfc344b2e</guid>
            <category><![CDATA[agentic-infrastructure]]></category>
            <category><![CDATA[small-model]]></category>
            <category><![CDATA[agent-substrate]]></category>
            <category><![CDATA[sandbox]]></category>
            <category><![CDATA[agentic-software-factory]]></category>
            <dc:creator><![CDATA[Oleg Verhovsky]]></dc:creator>
            <pubDate>Thu, 17 Sep 2026 22:17:33 GMT</pubDate>
            <atom:updated>2026-09-17T22:29:57.910Z</atom:updated>
            <content:encoded><![CDATA[<p><em>We are automating software development, and quietly creating an entirely new production environment underneath it.</em></p><h3>TL;DR</h3><ul><li>Software factories promise <strong>ticket in → working code out</strong>. That is close to zero-touch for the developer, and fully hands-on for the platform team.</li><li>Below the waterline sits a new production environment: sandboxes, model gateways, identity, token budgets, context, security, evals, observability and capacity.</li><li>Our working thesis counts <strong>14 new infrastructure primitives and 62 operational tasks</strong>, 61 of them recurring. For a 200-developer organization, that is roughly <strong>431 person-days of setup and about 100 person-days a month</strong>, or 5.6 full-time engineers.</li><li>Cost is the measurable problem. <strong>Control is the bigger one</strong>, and with thousands of agents, manual operation becomes part of the risk.</li><li><strong>Google just shipped </strong><a href="https://cloud.google.com/blog/products/containers-kubernetes/agent-substrate-available-on-gke"><strong>Agent Substrate</strong></a><strong> on GKE</strong>: an open-source agent runtime built to run millions of sandboxes, with sub-second suspend and resume. When a hyperscaler builds a dedicated execution layer for agents, agent infrastructure has become its own category.</li><li>But Substrate solves <strong>one layer of the iceberg: execution</strong>. Model routing, budgets, identity, context, evals and policy still sit with the platform team.</li><li>The answer is not more humans. It is a control plane: <strong>intent in → infrastructure operates itself.</strong></li></ul><p>For the last two years, most of the conversation around AI and software development has focused on one question: <em>how much of the developer’s work can an AI agent automate?</em></p><p>That question is quickly becoming outdated.</p><p>The next phase is not one developer working with one coding assistant. It is hundreds, eventually thousands, of agents working in parallel. Agents writing features, fixing bugs, generating tests, reviewing code, updating dependencies, running security checks, modernizing old applications, preparing releases and responding to incidents.</p><p>We are moving from AI-assisted development toward software factories. And there is a large infrastructure problem hiding underneath them.</p><h3>The Agentic SDLC has an iceberg</h3><p>The promise of the software factory is beautifully simple:</p><blockquote><em>Ticket in → working code out.</em></blockquote><p>From the developer’s perspective, this can eventually become close to zero-touch. But removing the developer from many execution steps doesn’t remove the infrastructure required to execute those steps. It moves the work somewhere else.</p><p>An agent needs somewhere to run. It needs a sandbox containing the correct tools and dependencies. It needs access to repositories, package registries, APIs and sometimes production systems. It needs models, context, credentials and budgets. And somebody needs to secure it, observe it, update it and decide what happens when it fails.</p><p>That is the Agentic SDLC Infrastructure Iceberg.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cnwkpb2g6Ow7mei75zbM_Q.png" /></figure><h3>The software factory becomes a mission-critical system</h3><p><em>We are automating software development, and the software factory is quietly becoming a mission-critical system that somebody has to operate.</em></p><p>Imagine an organization running 500 agents simultaneously. These aren’t simply 500 API requests.</p><ul><li>One agent might execute for 30 seconds.</li><li>Another might spend two hours modifying a large repository.</li><li>Another needs a browser. Another needs Docker.</li><li>Another needs access to hundreds of repositories.</li><li>A release agent might need CI/CD credentials.</li><li>An SRE agent might need read access to production telemetry.</li><li>A modernization agent could run for days.</li></ul><p>Our thesis identified at least 11 types of agent workloads with very different runtime, model, access and infrastructure requirements. Now multiply that by:</p><p><strong>Teams × repositories × agent types × models × tools × environments.</strong></p><p>The result starts looking less like a developer tool and more like another production environment. And somebody has to operate it.</p><h3>What it actually means to operate hundreds of agents</h3><p>This is the part that is easy to underestimate. Once agents become production workloads, someone needs to:</p><ul><li><strong>Provision and operate sandboxes.</strong> Choose isolation technology, run clusters, configure autoscaling, warm pools, snapshots, garbage collection and capacity.</li><li><strong>Maintain environments for every repository.</strong> Build sandbox images containing the correct compilers, dependencies and tools, then continuously patch and refresh them.</li><li><strong>Operate the model layer.</strong> Manage providers, API keys, rate limits, gateways, routing policies, regional requirements and provider failover.</li><li><strong>Continuously handle model changes.</strong> Evaluate new model releases, migrate deprecated models and safely roll out new routes without degrading quality.</li><li><strong>Control AI spend.</strong> Define budgets by organization, team, project and agent; attribute every task; detect runaway loops; optimize sandbox and model costs.</li><li><strong>Manage agent identity and permissions.</strong> Give each agent the minimum access required, issue short-lived credentials and ensure secrets never reach model context.</li><li><strong>Secure agent execution.</strong> Control commands, network access, tools and MCP servers; defend against prompt injection and data exfiltration; protect the software supply chain.</li><li><strong>Keep context trustworthy.</strong> Maintain code indexes, documentation and ticket integrations, memory, skills and playbooks, while continuously detecting stale context.</li><li><strong>Continuously verify quality.</strong> Maintain eval suites and run regressions whenever models, prompts, context, routing or agent harnesses change.</li><li><strong>Operate fine-tuning where needed.</strong> Collect and clean training data, execute tuning jobs, manage adapters and safely roll them out or back.</li><li><strong>Observe the entire agent fleet.</strong> Trace sessions, measure success rate, queue time and cost, detect stuck or looping agents and maintain SLOs.</li><li><strong>Run incident response.</strong> Investigate agent-caused failures, recover failed sessions, handle provider outages and maintain disaster recovery.</li><li><strong>Schedule competing workloads.</strong> Prioritize agent types, teams and tasks while balancing latency, compute availability and cost.</li><li><strong>Integrate everything with the existing SDLC.</strong> Source control, CI/CD, ticketing, chat, branch protection, merge policies and the additional CI load created by agents.</li><li><strong>Continuously onboard teams and repositories.</strong> Configure environments, policies, tools and permissions as adoption spreads.</li><li><strong>Govern autonomy.</strong> Decide which agents can read, write, execute, merge or deploy, and when a human must remain in the loop.</li></ul><p>Almost none of this is “set it up once and forget it.” In our thesis, 61 of the 62 identified operational tasks carry recurring work. Eight are classified as very-high complexity and another 28 as high complexity.</p><p>This is the iceberg.</p><h3>Why this gets expensive surprisingly fast</h3><p>We tried to quantify this hidden infrastructure effort. Our thesis identified <strong>14 new infrastructure primitives</strong> and <strong>62 operational tasks</strong> around Agentic SDLC.</p><p>For a reference organization with 200 developers and five to six agent types, we estimate approximately:</p><ul><li><strong>431 person-days</strong> of initial setup</li><li><strong>~100 person-days every month</strong> to keep it running</li><li>roughly <strong>5.6 full-time engineers</strong> before anyone starts improving the platform</li></ul><p><em>These are our estimates, not industry benchmarks. Actual numbers will vary significantly by organization and architecture.</em> But the direction matters.</p><p>And this isn’t simply a DevOps problem. The work spans Kubernetes and platform engineering, LLMOps, security, context and developer-experience engineering, and FinOps.</p><p>This creates an interesting inversion:</p><blockquote><em>The visible cost of producing software may go down while the hidden cost of operating the software factory goes up.</em></blockquote><p>Every new agent type, model, repository and team adds another combination that somebody needs to operate.</p><h3>Cost isn’t actually the biggest problem</h3><p>Cost is measurable. The bigger concern is control.</p><p>We are giving increasingly autonomous systems the ability to write code, execute commands, access repositories, call internal tools and potentially interact with production environments.</p><ul><li>A stale sandbox can produce bad code.</li><li>Stale context can produce code that looks reasonable but is wrong.</li><li>The wrong model can waste significant money.</li><li>A runaway agent can consume resources indefinitely.</li><li>An overly permissive agent can expose credentials or sensitive data.</li><li>A compromised tool or MCP server can dramatically increase the blast radius.</li><li>And agent-written code itself becomes part of the software supply chain.</li></ul><p>The thesis therefore includes controls for short-lived credentials, secret isolation, tool policies, prompt-injection defenses, software-supply-chain security and complete audit trails.</p><blockquote><em>With ten agents, humans can compensate. With hundreds, it becomes difficult. With thousands, manual operation becomes part of the risk itself.</em></blockquote><h3>Google is already building below the waterline</h3><p>One reason I believe this transition is happening quickly is that major infrastructure vendors are already starting to address it.</p><p>Google’s <a href="https://cloud.google.com/blog/products/containers-kubernetes/bringing-you-agent-sandbox-on-gke-and-agent-substrate">Agent Sandbox on GKE</a> provides isolated execution environments for AI agents. Its <a href="https://cloud.google.com/blog/products/containers-kubernetes/agent-substrate-available-on-gke">Agent Substrate</a> work pushes further, toward infrastructure designed around the characteristics of agent workloads rather than treating agents exactly like traditional services.</p><p>I think this is an important signal. <strong>Agent infrastructure is starting to become its own infrastructure category.</strong></p><p>This isn’t an argument that Google, or anyone else, has solved the entire problem. Quite the opposite. Sandbox execution addresses an important piece of the iceberg. But look below the waterline again. Organizations still need to continuously decide:</p><ul><li>Where should this agent run, and how much compute does it need?</li><li>Which model should it use?</li><li>What context should it receive?</li><li>Which tools can it access?</li><li>Which credentials should exist for this task?</li><li>How much should this task be allowed to cost?</li><li>Is its behavior normal? Is its output trustworthy?</li><li>Should what we learned from this execution change how the next thousand agents run?</li></ul><p>The problem is bigger than sandboxing.</p><h3>We shouldn’t solve this by adding more humans</h3><p>The natural response to a new infrastructure layer is to create another platform team. That works initially. I’m not convinced it works at the scale toward which Agentic SDLC is heading.</p><p>Imagine hundreds or thousands of agents constantly starting, stopping and changing their behavior. Humans cannot continuously tune sandbox sizes, model routes, context, budgets, permissions, policies and capacity for every workload. And perhaps they shouldn’t have to.</p><p>If software development itself becomes increasingly autonomous, then the infrastructure supporting it should become increasingly autonomous too. Instead of humans operating every knob, the infrastructure should increasingly:</p><blockquote><em>Observe → Decide → Execute → Evaluate → Learn → Repeat</em></blockquote><p>Observe the workload. Understand the task. Choose an appropriate environment. Allocate resources. Select the model. Apply security policy. Watch execution. Measure quality, cost and performance. Then learn from the outcome.</p><p>Humans still define the important boundaries: <strong>intent, policy, security, business constraints and desired outcomes.</strong> But humans shouldn’t have to manually operate the complexity underneath them.</p><h3>Local models may become part of the infrastructure</h3><p>There is another interesting consequence. We don’t necessarily need a frontier model to make every infrastructure decision. Many operational decisions are narrower:</p><ul><li>What type of workload is this?</li><li>Which sandbox profile does it need?</li><li>Which model class is appropriate?</li><li>Is this task unusual? Is this behavior anomalous?</li><li>Should this workload be suspended?</li><li>Is this context stale?</li><li>Did this configuration produce a better outcome than the previous one?</li></ul><p>Smaller, specialized and potentially local models could eventually handle many of these decisions. That creates an interesting architecture: <strong>frontier models perform the difficult development work, and infrastructure models continuously optimize how that work runs.</strong></p><p>Because the infrastructure observes thousands or millions of executions, it can potentially become self-improving. Not autonomous in the sense of unlimited authority. Autonomous in the infrastructure sense: continuously reconciling the environment toward the organization’s desired outcomes, within human-defined policies and boundaries.</p><h3>We’ve seen this movie before</h3><p>Containers dramatically simplified packaging software. They didn’t eliminate infrastructure complexity. Once organizations started running thousands of containers, manually operating them became impossible. We needed another abstraction.</p><p>Kubernetes didn’t make infrastructure disappear. It created a control plane that continuously reconciled what humans wanted with what was actually running.</p><p>Agent infrastructure may be approaching a similar moment. Google’s work is a signal that even the execution layer needs to evolve for agentic workloads. But execution is only the beginning. The bigger question is:</p><blockquote><em>Who operates the entire iceberg when hundreds or thousands of agents are running simultaneously?</em></blockquote><p>I don’t think the answer can be another collection of dashboards, scripts, tickets and YAML. And I don’t think the answer can simply be more people.</p><h3>The next control plane</h3><p>The first generation of Agentic SDLC is focused on automating what developers do. The next challenge will be automating what is required to operate all those agents: sandboxes, compute, models, context, identity, security, cost, quality, observability, capacity and policies. And eventually, the continuous optimization of all of them together.</p><p>The software-factory promise is beautifully simple:</p><blockquote><em>Ticket in → software out.</em></blockquote><p>Perhaps the infrastructure promise needs to become equally simple:</p><blockquote><em>Intent in → infrastructure operates itself.</em></blockquote><p>The software factory automated the developer and left the platform team with a new production system to operate. That’s the central problem below the waterline.</p><p>The Agentic SDLC is here. The next infrastructure control plane may be what allows us to operate it at scale.</p><p><em>This article is based on my working thesis on Agentic SDLC infrastructure. The task counts, engineering effort and cost estimates are modeling assumptions rather than industry benchmarks. I would be very interested in comparing them with what teams running large agent fleets are seeing in production.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4fdbfc344b2e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>