<?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 Toni Bergholm on Medium]]></title>
        <description><![CDATA[Stories by Toni Bergholm on Medium]]></description>
        <link>https://medium.com/@tonibergholm?source=rss-aaeb752ae8e9------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*ruCZ_CT0wwh75kqe.jpg</url>
            <title>Stories by Toni Bergholm on Medium</title>
            <link>https://medium.com/@tonibergholm?source=rss-aaeb752ae8e9------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 01:44:49 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@tonibergholm/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 Tooling Under Software Still Assumes a Human Is in the Loop]]></title>
            <link>https://medium.com/@tonibergholm/the-tooling-under-software-still-assumes-a-human-is-in-the-loop-52e251f88fda?source=rss-aaeb752ae8e9------2</link>
            <guid isPermaLink="false">https://medium.com/p/52e251f88fda</guid>
            <category><![CDATA[ai-agents-in-action]]></category>
            <category><![CDATA[npm]]></category>
            <category><![CDATA[git]]></category>
            <dc:creator><![CDATA[Toni Bergholm]]></dc:creator>
            <pubDate>Fri, 10 Jul 2026 19:41:59 GMT</pubDate>
            <atom:updated>2026-07-10T19:41:59.723Z</atom:updated>
            <content:encoded><![CDATA[<p>Software development has changed faster than the infrastructure underneath it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*szF7tXXbFX3da-yzs_mO_A.png" /></figure><p>Agents can now generate code, run tests, open pull requests, install dependencies, and operate across several tasks in parallel. But the tools they rely on were designed around a very different assumption: one human developer, reading what they install, working in one checkout, and reverting code when something breaks.</p><p>That assumption no longer holds.</p><p>This is the idea behind [git.agentic](<a href="https://git-agentic.com/">https://git-agentic.com/</a>): open infrastructure for the parts of software development that break when agents become first-class participants.</p><p>Before going further, I want to thank **[Theo Browne](<a href="https://www.youtube.com/@t3dotgg)**">https://www.youtube.com/@t3dotgg)**</a>. His video, **[“I don’t have time to build these things, will you?”](<a href="https://www.youtube.com/watch?v=wEAb0x3wTRc)**">https://www.youtube.com/watch?v=wEAb0x3wTRc)**</a>, pitched two of the problems that became central to this work: an agent-auditable package registry and source control designed for fleets of agents. Theo put the challenge into the world plainly. I decided to build.</p><p>## The bottleneck moved</p><p>For most of software history, code was expensive to produce. The ecosystem optimized around that scarcity. Package releases were relatively deliberate. Repository operations assumed human-scale concurrency. Production behavior was treated as a consequence of the code currently checked into Git.</p><p>Agents invert that model.</p><p>Code generation becomes cheap, while verification, coordination, security, and recovery become expensive. An agent can install a malicious package before a person has even seen its name. Ten agents can need ten isolated worktrees immediately. A production agent can regress because its prompt, model configuration, memory, tool surface, or database schema changed — even when the application code is correct.</p><p>The bottleneck is no longer typing the code. It is trusting, operating, and reversing what the system does.</p><p>That is infrastructure work.</p><p>## Sentinel: inspect the package before it executes</p><p>The npm ecosystem assumes that someone understands what `npm install` is about to fetch and execute. Coding agents frequently do not. They encounter setup instructions, invoke a package manager, and execute lifecycle scripts with the ambient authority of the developer machine or CI environment.</p><p>[Sentinel](<a href="https://git-agentic.com/sentinel">https://git-agentic.com/sentinel</a>) is an agent-auditable security layer for npm.</p><p>It runs as a transparent proxy in front of the existing npm registry. Packages still resolve through normal npm workflows, but every tarball is inspected before install-time code runs. A deterministic audit engine produces a score and verdict, looking for signals such as obfuscation, secret access, unexpected network egress, lifecycle scripts, typosquatting, dependency confusion, maintainer changes, and suspicious release behavior.</p><p>When code does need to run, Sentinel applies a deny-by-default capability sandbox. Network access, home-directory reads, credentials, and filesystem writes are withheld unless policy explicitly permits them. An agent can request an exception, but only a human can grant one.</p><p>That boundary matters. “Human in the loop” should not mean a person watches every command scroll by. It should mean the infrastructure knows which decisions require human authority and makes those decisions impossible to bypass.</p><p>Sentinel deliberately begins as a compatibility layer, not a new package universe. Replacing npm outright would inherit its hardest problem on day one: network effects. A proxy can improve the security model now while preserving existing packages, lockfiles, and workflows.</p><p>The longer-term opportunity is larger: package distribution with native permissions, verifiable ownership, safer revocation, signed provenance, and interfaces designed for machines to evaluate rather than blindly trust.</p><p>## src-control: one checkout is no longer enough</p><p>Git is excellent infrastructure for human collaboration. But a traditional checkout still assumes one pair of hands, one directory, and one disk-backed working tree.</p><p>Agent fleets break all three assumptions.</p><p>[src-control](<a href="https://git-agentic.com/src-control">https://git-agentic.com/src-control</a>) explores a source-control model built around immutable snapshots and copy-on-write virtual worktrees. An orchestrator can fork several isolated worktrees in memory, let agents explore alternatives in parallel, and discard those experiments without leaving a trail of directories behind.</p><p>The repository state can also contain encrypted secrets and recipient-protected paths. Signed commits establish who produced a change, while sealed agent-session transcripts add the missing question: what process and intent produced it?</p><p>The important practical decision is that src-control works with Git. It can import existing repositories, export history back to Git, and synchronize with Git remotes. The goal is not to demand that the ecosystem migrate before the idea becomes useful. The goal is to introduce agent-native primitives at the point where today’s workflow starts to fail.</p><p>There are honest boundaries. Arbitrary compilers and test runners still expect filesystem paths, so those operations require temporary materialization. The cryptographic implementation also needs independent review before anyone should trust it with production secrets. This is working pre-1.0 infrastructure, not a declaration that decades of source-control engineering are finished.</p><p>But it demonstrates the core idea: worktrees can be cheap, ephemeral, parallel, attributable, and designed around agents rather than retrofitted for them.</p><p>## git.agentic: Git cannot roll back behavior</p><p>The third project begins where source control ends.</p><p>When a conventional service regresses, an engineer can revert its code. When a stateful agent regresses, the code may be only one part of the change.</p><p>An agent’s behavior is determined by a tuple: code, prompts, tools, model configuration, memory, and schema. Reverting only the Git commit can leave the prompt changed, the memory contaminated, the tool contract drifted, or the database migrated forward.</p><p>[git.agentic](<a href="https://git-agentic.com/git-agentic">https://git-agentic.com/git-agentic</a>) versions that whole tuple atomically. A commit records all six dimensions. A rollback restores them coherently, including memory state and reverse schema migrations.</p><p>This is not another observability dashboard. It is a recovery primitive. The test is simple: when an agent starts behaving badly in production, can the operator return the complete system to a known-good behavioral state with one operation?</p><p>The initial implementation focuses narrowly on stateful LangGraph agents with Postgres and pgvector. That constraint is intentional. Infrastructure becomes credible through one complete path, measured under real failures, before it earns a broad abstraction.</p><p>## Build the wedge, preserve compatibility</p><p>There is an easy way to make infrastructure projects sound visionary: announce replacements for Git, npm, and the rest of the developer stack.</p><p>There is a harder and more useful way: identify the smallest primitive the current stack cannot express, build it, and integrate it with everything developers already use.</p><p>That is the strategy across these projects:</p><p>- Sentinel sits in front of npm rather than waiting for a new registry ecosystem.<br>- src-control interoperates with Git rather than requiring a flag-day migration.<br>- git.agentic uses Git for code while extending versioning to the rest of agent behavior.</p><p>The ambition is foundational, but the adoption path is incremental.</p><p>## What comes next</p><p>The code is only the first proof.</p><p>The next phase is external validation: published releases, reproducible benchmarks, design partners, independent security review, measured false-positive rates, and real incidents where these tools prevent damage or turn hours of recovery into one command.</p><p>If agents are going to produce and operate a meaningful share of the world’s software, they cannot continue borrowing infrastructure whose safety model depends on a human reading every line and approving every action.</p><p>We need package security that agents can query, workspaces they can fork without polluting the machine, provenance that records how a change came to exist, and rollback that restores behavior rather than code alone.</p><p>That is what I am building at **[git-agentic.com](<a href="https://git-agentic.com/)**">https://git-agentic.com/)**</a>. All three projects are open source under Apache 2.0:</p><p>- [git.agentic on GitHub](<a href="https://github.com/git-agentic/git.agentic">https://github.com/git-agentic/git.agentic</a>)<br>- [src-control on GitHub](<a href="https://github.com/git-agentic/src-control">https://github.com/git-agentic/src-control</a>)<br>- [Sentinel on GitHub](<a href="https://github.com/git-agentic/pkg-registry">https://github.com/git-agentic/pkg-registry</a>)</p><p>And once again: thank you, **Theo Browne**, for pitching the problems clearly — and for ending with the right question.</p><p>Will you build them?</p><p>Yes.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=52e251f88fda" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>