The Method
An agentic SDLC that holds up in production.
Coding agents make individuals faster. Getting a whole team faster takes a working system: where humans spend attention, how work is sized, how it’s reviewed, and how every cycle makes the next one easier.
The Loop
Eight stages, sized to the work.
Small changes skip straight to building. Ambiguous or risky work gets a spec first. The lane is chosen up front and visible to everyone.
| Stage | What happens |
|---|---|
| Ground | Once per repo: map it, detect how to build and test it, and agree the non-negotiables. Output: a short AGENTS.md, a constitution and a working `verify` task. |
| Frame | Turn a ticket or a request into a goal, context, constraints and a definition of done. Pick the right lane. |
| Spec | For ambiguous or risky work (the Spec-first lane): agree the behaviour, constraints and acceptance criteria before planning. Other lanes skip it. |
| Plan | Explore read-only, then write a plan of small, independently reviewable slices with the tests that prove each one. |
| Build ↔ Verify | Implement one slice at a time and run the verify command until green. Escalate the model when stuck, not by default. |
| Review | An independent reviewer on a different model checks the diff against the plan, then your policy decides who else must approve. |
| Ship | Merge, watch CI, fix what breaks within a bounded number of attempts. |
| Compound | Turn what was learned into repository knowledge: solution notes, rules and tests the next run starts from. |
Principles
What we’ve seen actually work.
- Separate thinking from typing
- The highest-leverage human review is of the plan, before code exists.
- Give agents a check they can run
- One command that proves the repo is healthy turns a session you watch into one you can walk away from.
- Review in a fresh context
- The author never grades its own work. A different model family, reviewing the diff against the plan in a fresh, read-only context, catches different mistakes.
- Parallelize reading, not writing
- Fan out research freely; keep one writer per branch.
- Small batches
- Slices small enough to review in minutes get merged; giant agent PRs get stuck.
- Compound every cycle
- Every failure becomes a rule, a test or a note, so the same mistake doesn’t happen twice.
Autonomy
Earn autonomy, level by level.
Teams choose how far agents go on their own, per repository and path, and raise it as results prove out.
| Level | Agents may… |
|---|---|
| A0 | Suggest only |
| A1 | Edit and commit in the workspace; people push |
| A2 | Push and open pull requests; people merge |
| A3 | Merge once checks pass and independent review approves, with a human for runs that touched untrusted content |
| A4 | Operate fully autonomously for the classes of work you choose |
Measure
A scorecard, not vibes.
We measure outcomes against your own baseline: agent PR acceptance rate, first-review pass rate, time waiting for review, change failure rate, cost per merged PR and how often lessons compound, never lines of code.