The agent workflow harness

Agents say done. Canon checks the work.

You approve the plan before any code. Then a second agent with no memory of the build grades every acceptance criterion against the code, with the file and line as evidence, and a failing grade blocks the close. Plans, decisions and criteria stay in your repo as plain markdown. Run canon to open the Cockpit that runs it all.

MIT·Local-first·No account, no SaaS·Setup guide →
The Cockpit's Eval Report for a demo ticket, Make delete button confirm, graded by Opus. Each acceptance criterion has a pass and its evidence: the line of code, and a real Chrome run in which a click on Delete opened a confirm dialog and Cancel kept the items. The report starts with an evaluator run id and the commit it graded.
An evaluator's report on a demo ticket, graded by Opus: every criterion gets a pass and the line and run behind it.
  • evaluator-run-idProof a separate evaluator run wrote this report. sprint complete checks the run log for it.
  • graded-headThe commit that was graded. If the code moves on, the close blocks until a fresh evaluator grades it.
  • ModelThe grader's model is recorded in the report, and a report without the line can't close.
  • EvidenceA file:line and a run behind each pass. Here the evaluator also broke the code on purpose to check its test would notice.
See it work. Try one:
The Cockpit: the board, themes, the keyboard sheet, Admin with model tiers and the activity log, Projects, new tickets and a restricted terminal.
What it actually caught

Things a test suite can't reach.

A test that could not fail.

Across two sprints on a real agentic app, the gates never found a wrong number. They found a check that agreed with itself, safety branches nobody had ever executed, and screenshots timestamped 13 minutes before the commit that replaced the feature.

Each criterion gets a pass or fail with the file and line that back it, and nothing closes on a fail.

# both sides read the same list assert button.disabled == live_only # expectation written by hand offline_safe = {"question A", "question B"} assert button.disabled == (q not in offline_safe)
An evaluator report for a closed ticket: each acceptance criterion with a pass and the file and line that proves it.
930Evaluator reports in canon's own repo, Nov 5
1,407Tickets tracked in canon's own repo, Nov 5
3Agent CLIs in the cockpit
0Accounts or hosted services
A Cockpit session for a demo ticket: Claude Code running the command sprint start, loading the sprint skill, reading the repository and starting the ticket Add CSV support.
A real session: Claude Code starts a sprint on a demo ticket.
The daily loop

Open canon. Two commands. Everything lands in the repo.

Click a command · example output

Commands

terminal · your-project

      
in your repo

    Get started

    Plan. Build. Prove it.

    Install once, add the sprint skill to a project, then ask your agent to run sprint start.