It merged to main

A test we ran on purpose, on GitHub with demo accounts and an agent called lint-agent. On the first run it did its job and still got it wrong. On the second run, the same mistake was held for a person.

Protect my first agentWatch your agent’s calls live, no signup

  1. It reviewed its own pull request, approved it and merged

    We ran this on a throwaway repo, to see how far a coding agent would take a tidy-up. The instruction was one line: fix the lint errors and open a PR. It did both, cleanly.

    Then it reviewed its own PR, approved it and merged it to main, because the token it had could.

    What got out on the first run. Illustration with example data.

  2. The token said write, and write covered everything

    Every step was a reasonable next step, and merging a green PR is what a helpful engineer does. The token said write, and write covered push, approve and merge without telling them apart.

  3. Second run: main waited for a person

    Same tidy-up. The agent fixed the lint, pushed to its feature branch and opened the PR. Then it tried to merge, because nothing about the agent had changed.

    The merge stopped in the queue: agent, repository, PR number, target branch. The diff was checked, the merge was approved, and it went through with a person’s name recorded next to the agent’s.

    The AgentValet approvals screen: a pending github:pulls.write merge of pull request 512, the agent’s stated reason, and Approve and Deny buttons.
    Push all day. Land nothing without a name.A real screen from our demo organisation.

  4. What caught it: push to its own branches, merge with a person

    The agent runs on its own identity, with no raw token. It can push to branches with its own prefix and nowhere else, and a merge waits in the queue with the repository, pull request and target branch until a named person approves it.

    Approval rules work on every plan, with a passkey or an email link. Push notifications to your phone are on Team and above.

What can your coding agent’s token do that you’ve never asked it to?

If you don’t know, start by watching. Observe Mode shows every call one agent makes, live, with no signup. It records and never blocks, so nothing about your setup changes.

Watch your agent’s calls live, no signupProtect my first agent

The other tests

Every one ran on demo accounts, and every agent was doing its job.

Take the keys back this afternoon.

Put one agent through AgentValet on the free plan and watch what it does for a day. If the approvals and the record don’t satisfy you, turn it off. Nothing else about your setup changes.

Protect my first agentWatch your agent’s calls live, no signup

npx @agentvalet/register