05 from AI Agent Governance — One Engineer, Over a Hundred Repositories and the Decisions No Agent Makes

Decision rights for AI agents when nobody reviews

Steel block with three guarded switches, AI agent governance

Required reviews assume a second person, and required checks assume a rule in every repository that names the right job and stays current. A maintainer working with agents often has neither.

This procedure decides, before the agents start, which decisions stay with a person and which a check may make, and holds each one with configuration rather than trust.

  1. 1

    Write down what stays human

    Money, publication and anything that cannot be undone. The person whose account the agents commit under owns the list; the security lead signs it off.

  2. 2

    Hold history with one ruleset

    Every default branch gets the same ruleset, with the rules non_fast_forward and deletion, applied through gh api to repos/{owner}/{repo}/rulesets.

  3. 3

    Hold money with a budget action

    AWS Budgets attaches a deny-all policy to the role that creates resources once the month's actual cost crosses a ceiling; the EKS platform's bootstrap sets thirty dollars.

  4. 4

    Keep keys out of reach

    Model and cloud access both come from workload identity federation, so no key is stored to leak. A token reaches a command line on standard input from the keychain, never as an argument.

  5. 5

    Let the verification decide the merge

    Replace gh pr merge --auto with a workflow started by the verification's completion (on: workflow_run) that merges only when it concluded success, then starts the verification again on main, because a push made with GITHUB_TOKEN starts no workflows.

  6. 6

    State the cooldown

    Write cooldown: default-days: 3 into dependabot.yml, so shortening the three-day wait is a visible diff, and fail any repository that shortens it.

  7. 7

    Show every gate its violation

    Plant the change each check must refuse, run the plants daily, and fail the build when a check stays quiet or refuses for the wrong reason. Publish what still gets through.

Who decides what

DecisionHeld byOwner
Merge a dependency updateThe verification's own success on that commitPlatform team
SpendA budget action that attaches deny-all at the ceilingThe account owner
PublishA person's go-ahead; no pipeline publishes on its ownThe person named on the list
Rewrite or delete historyA ruleset: no force-push, no deletionThe repository owner
Shorten the cooldownA fleet rule that fails the repositoryPlatform team
Trust a checkA planted violation it has to nameSRE, with the security lead reading every miss

What shows it works

Each is a public file at a pinned version, or a page that counts it live.

The reproduction ran in a private test repository; the source line and the workflows are public. Holding decisions this way costs speed where speed is dangerous: an update opened while verification is broken waits for a person, and OpenSSF Scorecard's code-review check stays at zero.

Capability is the vendor's problem. Decision rights are yours.

The long version is the post AI Agent Governance. GitHub Docs, Automating Dependabot with GitHub Actions; GitHub Docs, pull request merge states; GitHub Changelog, default package cooldown, 14 July 2026.


Vladimir Mikhalev

Docker Captain  ·  CNCF Ambassador  ·  IBM Champion

The Verdict — production-tested analysis on YouTube.

The Verdict

Inconvenient truths about shipping in the AI era

Container security, platform engineering, and the agentic shift — tested in production, argued without the hype. The verdict reaches your inbox the moment there's one worth sending.

The other blueprints

Decision rights for AI agents when nobody reviews
https://heyvaldemar.com/blueprints/agent-decision-rights/
Author
Vladimir Mikhalev
Published
2026-11-13
License
CC BY-NC-SA 4.0