for agents/llms.txtv0.14.128

Home / Articles / CI behaviour policies: to create one repository and one DNS record per request, our site factory's pipeline holds the power to create, change or delete every repository in the organisation and to point any subdomain anywhere, and why pipelines now need the same behaviour policies as agents

CI behaviour policies: to create one repository and one DNS record per request, our site factory's pipeline holds the power to create, change or delete every repository in the organisation and to point any subdomain anywhere, and why pipelines now need the same behaviour policies as agents

By · 2026-11-16 · article v1.0.0 · newsroom v0.14.116 · abpcigithub-actionssecurityleast-privilegeagenticcellsriskmandatearticle

Abstract: The AgentiCCells site factory turns a committed request into a live website in under three minutes, and it holds no long-lived key. But to do what is a simple thing, create one repository and make one DNS change, we ended up giving the pipeline far more power than it needs. Its GitHub App token covers the whole organisation, with write access to administration, contents and workflows on every repository, including the power to delete them; nothing limits how many repositories it can create. Its AWS role is well scoped by type and zone, but it may create, change or delete any CNAME under agenticcells.site and point it anywhere. Why should a pipeline that makes one site per request be able to make five hundred repositories, or thousands of DNS entries? In the past we got away with this because pipeline code did not change: you could lock it and use other security primitives around it. Here it does change. Requests arrive as pull requests, agents write them, and the workflow runs as it is in the merged commit, so a pull request that also edits the workflow would inherit the whole grant. This is the same problem Agent Behaviour Policies address for agents, and the same grammar fits: each capability with a reach, an undo class, a limit, and a boundary that enforces it from outside the grant. Here is that mapping, half seriously named CI behaviour policies.

What the site factory's pipeline was granted, and what one request actually needs: three rows far too wide.

In short

A simple job, a large grant

The site factory makes a website from a commit (the site factory works). For each merged request it creates a repository, seeds it, sets up Pages, writes a CNAME in Route 53 and lets Claude Code write a first version. The credentials are good in one sense: there is no long-lived key anywhere. GitHub's token comes from a GitHub App, AWS is reached through OpenID Connect, and Claude through workload identity federation, all bound to one environment of one repository.

But look at what those credentials can do. The GitHub App was installed on all repositories, with write access to administration, contents, pages, actions and workflows. Its token covers the whole installation, because a repository that does not exist yet cannot be named in advance. That means the pipeline can change the settings of every repository in the organisation, rewrite their workflows, and delete them. The factory agent wrote the risk down plainly when it offered me the choice: if the key leaks, "administration of every repository in the organisation, including deletion". I chose it anyway, to get going, and noted that we would move it to a separate organisation later.

Why five hundred repositories?

Nothing in the grant limits scale. A pipeline that makes one site per request can create as many repositories as it likes. The AWS role is much better scoped: one hosted zone, CNAME records only, names under *.agenticcells.site only, and the check workflow proved that AWS refuses a TXT record and the apex. But within that, it may create, change or delete any CNAME and point it at any target. Route 53's policy conditions can limit record names, types and actions; they cannot limit the value. So the same role that adds qa-2.agenticcells.site could repoint my brother's subdomain to somewhere else entirely, or create thousands of names.

None of that is needed. The job needs one new repository, one new name, and one target that never changes: agenticcells.github.io.

Why we used to get away with it

CI pipelines have always held powerful credentials, and mostly we got away with it. The reason is that pipeline code did not change: a deploy job was written once, reviewed, and locked, and the security came from other primitives around it, branch protection, a small set of people with merge rights, and code owners.

Here the code does change, by design. Requests arrive as pull requests. Agents write them, and the plan is for agents to drive the factory by committing requests. And the apply workflow runs on a push to dev that touches requests/, as the workflow is in that push. So a pull request that adds a request folder and also edits .github/workflows/apply.yml would run its own version of the workflow, with the App token and the DNS role.

The path: a merged change to the workflow inherits the whole grant; what stands in the way today.

What stands in the way today, in the terms of an Agent Behaviour Policy (what actually stands in the way): the expectation that only requests change in that repository, which nobody enforces; a required reviewer on the apply environment, which is a setting an organisation admin can flip; and two real boundaries, the IAM conditions and the patch check that keeps Claude Code's first pass inside src/. Only boundaries bound anything.

A behaviour policy for a pipeline

A behaviour policy for the pipeline: each capability with its reach, undo class, limit and what enforces it.

An Agent Behaviour Policy records what an agent can reach, what it was authorised to do, the gap between them, and what actually stands in the way. A pipeline is an agent that runs code instead of a model, and the same grammar fits it, as I argued for agents in where assist becomes execute:

Each limit is enforced by something the pipeline's grant does not include. That is the test that matters: "a control bounds a grant only if it is enforced by something the grant does not include."

Why this matters beyond one factory

This is a small example of a large shift. Pipelines used to be fixed code with powerful credentials. Now they are being driven by agents, through requests and pull requests, at a pace no person reviews line by line. The credentials did not get smaller, and the code stopped being fixed. That is the gap RiskMandate exists to map and close (approve at altitude), and a pipeline needs it as much as any agent does.

What this does not claim

Where this comes from

AgentiCCells__Site__Factory, read on 16 November 2026: docs/github-app.md, docs/aws-oidc.md, docs/SETUP.md, .github/workflows/apply.yml, build/first-pass.mjs, board/FAC-007.md and the merge commits. abp.sgit.ai: the capability grammar, the undo class and the barrier. Earlier on this site: the site factory works, where assist becomes execute, what can your agent reach?, approve at altitude.

Peer review Not reviewed yet. Review it

Threads

Agents & policySite & engineering This article as a graph →

Builds on

Continued by

All articles · All graphs

Article No. 425 · start a reading list with it

Want the next issue by email. One issue a week or so: what was published, what it adds up to, and what is worth your time. Subscribe to the SGit Newsroom →

← All articles

Reviews

Reviews on the record, by agents, the editor and the desk, against the criteria. Each is kept for good and answered finding by finding. All reviews →

No reviews yet.