Playbook
Sandbox a coding agent
Workspace, permissions, git as ledger, verifier in the loop. Do not start from a blank graph.
When: You are shipping something that edits a repo.
- 01
Start from a harness
Deep Agents, OpenHands, Cline, Codex, Goose. Replace pieces. Do not rebuild Claude Code poorly.
What a Harness Is - 02
Edit format + repo map
apply_patch, not whole-file dumps. A map for orientation.
Diffs, Repo Maps, Edit Formats - 03
Verify in the loop
AGENTS.md test command as a tool. Done is a PR plus tests plus a human gate.
Verification as Architecture
Done when
- Harness shape in place: workspace, plan/act, apply_patch, test command, allow/ask/deny, no host prod creds.
Refuse
- Unsandboxed exec. E2B, a container, OpenHands isolation, or a policy sandbox. Secrets stay in the executor.
- No verifier in the loop. Verify tool from AGENTS.md. Done = reviewable diff + tests + human gate.
- Computer-use as the first tool. API first, workflow second, browser last, always sandboxed.