How it runs
One conversation, then hands off.
- 01
Talk
Describe the feature. The agent pushes back on anything that clashes with your code or your stack.
- 02
Confirm
You mark each requirement ready. Confirming is the handover.
- 03
Merged
Design, steps, tests and the pull request run without you.
- Requirementsyou + opus
- Designopus
- Step 01 · export modelsonnet
- Step 02 · csv writersonnet
- Step 03 · download routesonnet
- Acceptance testssonnet · blind
- Final validationcode
- Pull requestgithub
- 00:31 requirements.md confirmed · 6 of 6 ready
- 01:33 handover · worktree ternolab/014-invoice-export
- 02:36 design: reading context and 9 specs
- 06:14 decision 001 · reuse the existing Invoice query
- 08:03 decision 002 · stream rows, no temp file
- 09:52 design: 3 steps written
- 10:55 step 01: tests first, then code
- 14:33 step 01: 4 tests pass
- 15:20 step 02: tests first, then code
- 19:14 step 02: 1 of 5 tests fail · quoting
- 20:48 step 02: retry 1 of 1
- 24:10 step 02: 5 tests pass
- 24:57 step 03: tests first, then code
- 27:02 acceptance: written from requirements only
- 29:07 step 03: 3 tests pass
- 30:56 acceptance: 6 tests, one per requirement
- 31:58 validate: build · typecheck · 212 tests
- 36:23 validate: all checks pass
- 37:26 pr #48 opened · waiting for CI
- 42:07 pr #48 merged into main
Built to run without you
1Tests written blind
The acceptance tests come from the requirements alone. The agent that writes them never sees the code.
- requirementdecisionstep
2Every guess on record
Each choice an agent makes is a file in a graph, linked to the requirement it touches. It merges with the code.
3Stops when it should
A failed step retries once. Then the design agent rewrites its prompt, up to three times. Then it comes back to you, with questions.
4Your repo, your CI
Mill opens a pull request and merges it when your checks pass. You review once per repo: the onboarding pull request.
Spec plate
What does the work.
Large models where the thinking is. Small models where the typing is. Deterministic code wherever a check can be written down.
- Requirements, design
- Opus
- Steps, acceptance tests
- Sonnet
- Checks
- Plain code, not an agent
- Merge
- GitHub PR or local git
- Features at once
- No limit
- Human reviews
- One per repo
TL-MILL · SERIES 1
- 01
AI opportunity assessment
Where AI is worth applying in your business, and where it is not.
- 02
Pilot build
One workflow taken from brief to production, with an evaluation suite that establishes whether it improved on the baseline and reports when it begins to drift.
- 03
Secure AI design
Architecture and threat modelling for systems that handle company information or act on it.
- 04
Customer security diligence
The AI section of the questionnaire holding up a contract, answered accurately, evidenced, and documented so that the next one costs a fraction of the first.
- 05
Fractional AI lead
Continuing technical direction for teams past their first project.