← back to lab
lab / background agents / 2026-06-05

Background agents need handoff shelves

A background coding agent can be genuinely useful: it can pick up a small issue while a human works elsewhere. But “it opened a PR” is not enough evidence. The useful question is whether the trigger, workbench, tool marks, rejected parts, and return shelf are visible before review.

Current handoff claim

“Mention the agent and it opens a pull request later.”

Check only public docs, repo behavior, issue/PR traces, or a synthetic fixture. A passing result means one bounded trial, not autonomous engineering trust.

The eight shelves

Trigger
Name which surfaces can start work and who may use each one.
Issue binding
Bind the run to repository, branch, files, acceptance criteria, and scope.
Sandbox
State what is isolated: filesystem, network, packages, account access, browser state, or only the process.
PR / fixup
Show where code lands and how authorship, comments, and evidence are labeled.
Review cut
Mark exactly where autonomy stops before merge, external calls, or protected changes.
Failure signal
Leave a trace for crashes, failed tests, irrelevant patches, and incomplete runs.
Recovery
Make cancel, revert, resume, and safe-point behavior visible.
Claim size
Separate helper bot, background developer, and self-running team claims.

Source door

This gate was sharpened by public descriptions of background coding-agent handoff systems. This page does not endorse, install, execute, or copy any specific system. It keeps the reusable test: can the handoff be inspected and recovered before a human is asked to approve the result?

Stop rule

If trigger, issue binding, sandbox, review cut, failure signal, and recovery are not visible, keep the system as inspiration only. The next step is a smaller public-doc receipt or synthetic fixture, not an unsupervised coding loop.