Skill · job · coding-agent
Coding agent
Ship a coding agent: AGENTS.md, apply_patch, repo map, tests in the loop, allow/ask/deny, human gate. Use for PR agents, SWE-bench-like work, repo edits, or when comparing Codex, OpenHands, Cline, Aider, Goose, Deep Agents. Start from a harness, not a blank graph.
GET /api/canon/skills/coding-agent?format=md
Every serious coding agent converges: workspace, skills, plan/act, diffs, sandbox, verifier, human gate. Computer-use is not the first tool.
When
The product is software engineering — PRs, tests, and a desk, not a chat wrapper.
Do
- 01
Start from a harness
Codex, OpenHands, Cline, Aider, Goose, Deep Agents. Replace pieces. Do not rebuild Claude Code poorly.
- 02
Edit format + repo map
apply_patch, not whole-file dumps. A map for orientation. Grep over read-whole-repo.
- 03
Verify in the loop
AGENTS.md test/typecheck/build as tools. Compact failures. Done is a reviewable PR + tests + human gate.
- 04
Gate writes
Plan mode denies writes. Act asks. Host prod creds never enter the sandbox.
Don't
- Unsandboxed exec on the laptop with prod AWS.
- Declare victory because the model said 'done'.
- Screenshot the IDE instead of using git and tests.
Hard rules
- Done = reviewable diff + tests + human gate.
- No verifier in the loop is a block.
- Skills for depth; AGENTS.md for commands and boundaries.
Refuse
- Unsandboxed exec — A shell or code-act tool on the host with production credentials.
- No verifier in the loop — The agent writes code and declares victory. Tests, typecheck, and build are optional.
- Unbounded loop — No max-turns, token, wall-clock, or money budget. The agent rambles until the bill or the context dies.
- Mega-tool — One function whose argument is a JSON blob of 'whatever'. An untyped shell with extra steps.
- Computer-use as the first tool — A browser or desktop agent used where an API, a workflow, or a grep would do.
Load with this
Load next
- Design a harnessThe product is a desk, not a chat wrapper. Claude Code, Codex, OpenHands, Cline, Goose, Deep Agents converge here.
- Write AGENTS.mdA coding agent will touch this repository. The constitution is short, true, and runnable.
- Plan / actSide effects need a human-agreed shape. Plan is a file, not a thought.
- Allow / ask / denyThe agent can touch a workspace, a browser, or production. The model proposes; code decides.
- Spec-driven workYou can score the work. The agent would otherwise wander.
Trigger tests
Should fire
- “Build a coding agent that opens PRs”
- “We want something like Cursor/Codex on our repo”
- “The agent writes code but never runs tests”
Should not
- “A support bot for refunds”
- “Chat with our PDFs”