<?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 Ravinder Singh on Medium]]></title>
        <description><![CDATA[Stories by Ravinder Singh on Medium]]></description>
        <link>https://medium.com/@singhravi1979?source=rss-fc15cb7317b0------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*QV2z3QSunTIAC8IP.</url>
            <title>Stories by Ravinder Singh on Medium</title>
            <link>https://medium.com/@singhravi1979?source=rss-fc15cb7317b0------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 07:42:39 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@singhravi1979/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[How to Build an Agentic Engineering Workforce Without Losing Control]]></title>
            <link>https://medium.com/@singhravi1979/how-to-build-an-agentic-engineering-workforce-without-losing-control-6a2e65c7cfe9?source=rss-fc15cb7317b0------2</link>
            <guid isPermaLink="false">https://medium.com/p/6a2e65c7cfe9</guid>
            <category><![CDATA[ai-sdlc]]></category>
            <category><![CDATA[ai-agents-in-action]]></category>
            <category><![CDATA[nanoclaw]]></category>
            <dc:creator><![CDATA[Ravinder Singh]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 11:05:09 GMT</pubDate>
            <atom:updated>2026-11-17T11:05:09.969Z</atom:updated>
            <content:encoded><![CDATA[<p>Container isolation, a credential gateway, and a human merge gate: what worked, what bit us, and a 30-day rollout for small teams.</p><p>What if an AI agent could take a ticket all the way to a green merge request, and still never see a single real credential or merge its own code? That’s the setup we piloted. This is what it takes to make it safe, and where it still hurts.</p><h3>The idea in one paragraph</h3><p>Treat each AI agent like a junior teammate with a job description, its own login, its own workspace and a manager who approves its plan. NanoClaw is the host that makes this practical: it runs each agent in its own container, connects it to chat (Teams, Slack and others), routes each conversation thread to its own session, and lets you hand it tools and skills. You supply the roles, the guardrails and the human review.</p><h3>What the setup looks like</h3><p>Everything runs on one host, a single VM. People talk to agents through the chat tools they already use, and agents reach your tools only through a credential gateway.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Sg0tEUHyjJv24KJHzpCZuw.png" /><figcaption>Each role runs in its own container. Agents hold placeholder tokens; the gateway adds real ones. A person always merges.</figcaption></figure><p>Key parts:</p><ul><li><strong>Agent group:</strong> a role (persona, skills, tools, memory folder, container image). One group per job.</li><li><strong>Template:</strong> a saved definition of a role. You can stamp a group from it and re-stamp when the role changes.</li><li><strong>Sessions:</strong> one per chat thread, so two tickets do not share a conversation.</li><li><strong>Credential gateway:</strong> the agent never sees a real token. It calls tools with a placeholder and the gateway swaps in the secret for approved hosts and paths only.</li><li><strong>Skills:</strong> markdown procedures the agent follows (your team’s “how we do a ticket”). You can reuse skills you already use with a coding assistant.</li></ul><h3>A first workforce in five steps</h3><ol><li><strong>Pick one boring, repeatable job.</strong> Example: “take a ticket, plan the fix, open a merge request that passes CI”. Not “run engineering”.</li><li><strong>Give the agent its own identities.</strong> A bot user in the issue tracker and a bot or group token in the Git host, each with the lowest role that works (for code: Developer, never Maintainer). Never reuse a person’s token.</li><li><strong>Put credentials behind the gateway.</strong> One secret per host/path, assigned to this agent only. Check that no other agent can read them.</li><li><strong>Write the workflow as a skill.</strong> The procedure, the rules, the message templates. Keep it short and testable (more on that below).</li><li><strong>Wire one chat channel, mention-only at first.</strong> Rehearse on a scratch repo and scratch tickets before the real group.</li></ol><h3>Guardrails that mattered</h3><p>Treat these as requirements, not nice-to-haves. Each line is a risk, followed by how we handled it.</p><ul><li><strong>Agent merges or pushes to a main branch.</strong> Enforce it in the Git host (protected branches; Developer cannot push or merge). Also state it in the skill. The host rule is the real backstop; the skill is the first layer.</li><li><strong>Work starts without a person’s approval.</strong> Plan first, then wait for the requester only. Record the plan and approver in the ticket. This is a prompt-level rule; a hard gate (separate read-only planner and write-enabled worker) is a stronger option.</li><li><strong>Instructions hidden in tickets, comments or pages.</strong> Treat all fetched text as untrusted. Approval by the requester is the main defence. Rehearse with a planted injection.</li><li><strong>Agent sees raw secrets.</strong> Credential gateway with placeholders. Never print env or tokens.</li><li><strong>Cross-conversation confusion.</strong> Messages copied between sessions of the same chat are background only, never a request or an approval. Approval must be the triggering message and name the plan version.</li><li><strong>Runaway cost or time.</strong> Per-ticket ceilings on turns and minutes. Cap concurrent tickets.</li><li><strong>Repo with weak branch protection.</strong> Audit protection before granting access. Fix the weak ones first.</li></ul><h3>Test the workflow like code</h3><ul><li>Run <strong>rehearsal scenarios</strong> on scratch repos: a normal ticket, an unclear ticket, a red pipeline, a repo with no pipeline, a manual-only pipeline, a ticket with a planted injection, an approval from the wrong person, two tickets at once.</li><li>Put template and helper checks in CI (message shape, banned commands, helper scripts).</li><li>Record results in a rehearsal log next to the skill.</li></ul><h3>Communication: make the agent legible</h3><p>People will not trust an agent that goes quiet. What worked:</p><ul><li>Acknowledge within a minute of the request.</li><li>Post the plan and say exactly how to approve, change or cancel it.</li><li>Acknowledge the approval at once, then post short milestone updates (cloned, tests passing, MR open, pipeline state).</li><li>Use a fixed shape for every message: what happened, what is needed from the reader, what happens next and roughly when.</li><li>Stop at a green merge request. A person reviews and merges.</li></ul><h3>Pros and cons as an AI-SDLC workforce</h3><h4>Pros</h4><ul><li><strong>Isolation by default.</strong> Each role runs in its own container with its own files and credentials.</li><li><strong>Safe credential handling.</strong> The gateway pattern keeps real tokens out of the agent.</li><li><strong>Roles as data.</strong> Templates make roles repeatable; a role change is a re-stamp, not a rebuild.</li><li><strong>Thread-per-session routing.</strong> Natural fit for ticket-per-conversation work.</li><li><strong>Chat-native.</strong> People use the tools they already have. No new UI to adopt.</li><li><strong>Reuses existing skills.</strong> If your team already writes procedures for a coding assistant, much of that carries over.</li><li><strong>Small footprint.</strong> A single mid-size VM ran the pilot, so a startup can begin cheaply.</li><li><strong>Human stays in the loop.</strong> The plan-approve-review shape is easy to explain to a team and to auditors.</li></ul><h4>Cons and limits we hit</h4><ul><li><strong>Guardrails are mostly prompt-level.</strong> The approval gate is an instruction the model follows, not a technical lock. Plan for the Git host to be your hard backstop.</li><li><strong>No per-role resource limits.</strong> Memory and CPU limits were host-wide, so one heavy role raises the cap for all. Sizing needs care.</li><li><strong>No built-in cap on concurrent sessions.</strong> You enforce the ticket limit yourself.</li><li><strong>Chat platform quirks.</strong> Channel messages may only reach the bot when it is mentioned unless you add the platform’s resource-specific consent permissions. App manifests have their own requirements (the upload fails without a web application id).</li><li><strong>Cross-session copies.</strong> NanoClaw copies messages into sibling sessions for context. That is useful, but your skill must say these copies never count as requests or approvals.</li><li><strong>Latency.</strong> Reading a ticket and a codebase and writing a plan took 5 to 18 minutes. Without interim messages it feels broken.</li><li><strong>Cost is hard to predict.</strong> Per-ticket cost needs telemetry you set up deliberately. Subscription logins may hit rate limits under bulk or concurrent use.</li><li><strong>Operational work you own.</strong> Image builds, deploys, restamps (only when idle), token expiry and rotation, branch-protection audits.</li><li><strong>Moving target.</strong> Local patches are a burden on upgrade, so prefer fixing things in skills and configuration.</li><li><strong>Repo variety.</strong> Old runtimes, Docker-in-Docker, databases and browsers in tests do not fit a simple image. Leave those parts to CI and say so in the MR.</li></ul><h4>When it fits</h4><ul><li>A small team with a repeatable ticket-to-MR workflow and strong CI.</li><li>You can enforce branch protection and want a human to merge.</li><li>You value auditability (plans in the ticket, approver recorded).</li></ul><h4>When to think twice</h4><ul><li>You need hard, technical separation between planning and acting from day one.</li><li>Your repos lack CI or branch protection and you cannot fix that first.</li><li>You need strict per-role resource limits or a vendor SLA.</li></ul><h3>A 30-day rollout for a startup</h3><ul><li><strong>Week 1:</strong> Host up, one role, bot identities, gateway, scratch repo. Rehearse the happy path.</li><li><strong>Week 2:</strong> Rehearse the failure cases and the injection test. Add CI checks for templates and helpers.</li><li><strong>Week 3:</strong> Real group, one channel, mention-only, one requester team. Limit to two concurrent tickets.</li><li><strong>Week 4:</strong> Review every MR, tune the skill and messages, measure cost per ticket, then decide on a second role.</li></ul><h3>Lessons in one list</h3><ol><li>Enforce safety in the Git host, then in the skill.</li><li>Approval means the requester’s own triggering message that names the plan version.</li><li>Silence reads as failure. Acknowledge, update, and say what you need.</li><li>Rehearse failure modes on scratch repos before the real channel.</li><li>Prefer configuration and skills over patching the platform.</li><li>Measure cost and time per ticket from the start.</li><li>Keep people as the merge gate.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6a2e65c7cfe9" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>