A practical checklist for supervising AI coding agents across branches, worktrees, reviews, approvals, and human intervention points.
AI coding agents -- Claude Code, Cursor, Codex, Aider, OpenCode, and similar tools -- are becoming powerful enough to do real work. They can plan, write, test, and iterate. Sessions that used to take minutes now run for hours. Agents touch more files, operate across more surfaces, and increasingly run in parallel.
When that happens, the bottleneck shifts.
Code generation is no longer the hard part. Supervision is.
The questions that matter become:
- Which agent is working on what right now?
- Who has reviewed what, and when?
- What is the approval state before anything merges?
- When does a human need to step in?
- What happened in the last session?
Most developers do not have good answers to these questions. This repo is an attempt to help with that.
Worktrees isolate code. They do not fully solve state, review, approvals, ownership, or human intervention.
Git worktrees and branches are the right foundation. But they only address one layer of the problem -- the code layer. Everything above that (who reviews, what gets approved, when to , how to audit) still requires deliberate process.
See worktrees-vs-control.md for a full breakdown. A minimal checklist for any agent session:
- Scope is defined before the agent starts
- Agent is working in an isolated branch or worktree
- Files the agent may touch are known in advance
- You know how to or interrupt the agent
- Diffs are reviewed before anything is staged or merged
- Tests pass and you understand what they cover
- No secrets, credentials, or config values were written by the agent
- Changes are approved by a human before they reach main
For the full version, see checklist.md.
| File | What it covers |
|---|---|
checklist.md |
failure-modes.md
`worktrees-vs-control.md`
`sample-agent-review-log.md`
CONTRIBUTING.md
Agent tooling is moving fast. Most of the discourse focuses on what agents can do. Less attention goes to what breaks when you actually run them at scale -- multiple sessions, multiple files, real production codebases.
This checklist is not a framework. It is a collection of practical controls that developers can adapt to whatever workflow they already use.
This checklist is maintained by Useful Softworks. We are also exploring AgentLeash, a local-first control layer for builders using AI coding agents. Private beta applications are open at https://agentleash.dev/
AgentLeash is a product of Useful Softworks, a software brand operated by Chain Bridge Labs LLC.
See CONTRIBUTING.md. Pull requests are welcome for new failure modes, improved checklist items, and tool-specific notes.
MIT. See LICENSE.