Stage 4: Build & validate

Concept — bounded agent execution and deterministic guardrails

You leave with — a remediated seam, built inside boundaries you set

AI generates inside tests, rules, gates, tool permissions and contract checks — not inside a conversation.

What each agent receives — and what it does not

   ✓  GIVEN                              ✗  NOT GIVEN
   ─────────                             ────────────
   the validated specification           the other repository
   its repository brief                  this action guide
   its implementation brief              your reasoning about the seam
   the acceptance criteria it owns       the other agent's return
   that repository's CLAUDE.md

One scoped implementation context per repository. The exclusions are the design, not an oversight.

The facilitator demonstrates — the shape of an implementation prompt

Watch how it states the outcome and the boundary, and then refuses to describe the fix.

A prompt that specifies the diff is just a slower, more expensive way of writing the diff yourself. You are buying the agent’s ability to find an implementation; if you hand it one, you have bought nothing and still have to review it.

▶ Your turn — brief the other repository

Adapt it to what that repository actually owns.

The two are not symmetric. One of them may correctly conclude it needs no production change at all — in which case its job is to say so and prove it.

That is a real outcome, not a failed one. Establishing that a repository is already correct, with evidence, is engineering work. Some of the best returns you get today will contain no diff.

Expected return from each agent

   Files changed         Verification run
   ACs addressed         Open concern
   Tests added           No-scope-expansion confirmation

Two things that will actually happen

⚠ An agent will offer to fix something in the other repository. Refuse it. That is the seam, and the seam is yours. An agent that can reach across the boundary has no way to know what it would break, because it cannot see the other side.

⚠ An agent will find nearby code that looks reusable and unfinished. It compiles. It shares infrastructure with the refund path. It looks abandoned mid-change.

Adjacency is not authorisation. If it is out of scope, the correct action is no diff plus a recorded reason.

Verify as you go

cd pgs-tta && mvn verify
cd ../pgs-payment-processor && mvn verify

   ╭──────────────────────────────────────────────────────────────╮
   │  Both green does not mean done.                              │
   │  That is the entire premise of this lab.                     │
   ╰──────────────────────────────────────────────────────────────╯

Writes follow the rollout order you decided in Stage 3.

⌘  /hand-off

---