§01 / Who's on the Team
6 agents · one entry eachWho's on the team.
Six agents make up the team — the same engineering team that builds and maintains every agent in this shop. Open a row for each one's spec: what we gave it, and where we made it stop.
Given 'build X', our Architect turns it into a plan the team can run without stepping on each other. It reads the target repo and the registry record first, then cuts the work into ordered, path-scoped tasks: who touches what, in what order, what runs in parallel, and what 'done' means.
We hold it to one build_task proposal per writer, and it prefers extending what is already there over inventing something new.
It writes no feature code, and it never pushes, merges, or deploys. Every plan ends with a QA task as the gate — never a build task — and its task order becomes the merge order later.
— from the build notes
Backend builds the server side of one assigned task, in its own git worktree, on its own branch — matron/<task>-backend — and we keep it in its lane: APIs, services, auth, DB schema, migrations, queries, pipelines.
We require migrations to go out additive and reversible, and it adds unit tests and verifies the code actually runs locally before handing the branch to Forge.
If it needs something outside its path scope — a shared type, a schema another task owns — it stops and reports back instead of reaching across. No push, no PR, no merge, and it never runs a migration against a real database.
— from the build notes
Frontend builds the UI slice of one task, in its own worktree, on matron/<task>-frontend — components, pages, client state, styling, accessibility, in whatever the repo already uses (Next, React, Tailwind). It reads the existing components, hooks, and Tailwind tokens and matches them.
We do not let it drag in a new state library or a UI kit the repo has never seen. It wires to the real API contract, and when the backend is not built yet it coordinates through Matron rather than guessing silently.
If it needs an API field that does not exist, it flags the gap so the tasks get sequenced right.
— from the build notes
Flow is our n8n engineer. It builds and edits workflow JSON, webhooks, node config, and the EmailProvider wiring behind the partner email-agent line, on matron/<task>-flow.
It validates offline before anything is real: the JSON parses, credentials and nodes exist, webhook paths do not collide, node ids stay stable. We treat one workflow JSON as one indivisible unit with one owner — it is never split.
Secrets never live in the file; credentials are referenced by n8n name or id, never pasted as tokens. Editing the exported JSON is normal file work — pushing it live is a separate deploy_workflow proposal you approve.
— from the build notes
QA is the last gate we put before your change becomes a PR. It runs the suite, type-check, lint, and build against every engineer's branch, then writes the integration and e2e tests on its own branch and actually exercises the work — runs the app, hits the endpoint, triggers the workflow.
You get one verdict: ship or no-ship, with the blocking defects named exactly. We hold it to Architect's acceptance criteria, not its own assumptions, and it captures real command output instead of guessing.
It does not fix feature code — a bug it finds goes back to Matron as a blocking defect for the engineer who owns it.
— from the build notes
Forge is the only agent we allow to merge. After the engineers build and QA says ship, it takes their branches and merges them in Architect's declared order onto one integration branch, rebuilding after each merge so breakage shows up early.
It resolves the trivial conflicts and drafts the PR body. Real conflicts or logic gaps go back to Matron — it integrates, it does not invent.
It never authors features, and opening the PR, pushing, merging to main, and deploying all stay separate proposals you approve. We also gave it a second job: scaffolding a minimal runnable skeleton when a POC gets the green light.
— from the build notes
Next step · At the counter
Heard the intro? Book a session — a trial on your own machine, a read-only Phase-0 assessment, or the full build. Nothing runs until you approve it.
§02 / Also on the Shelf
More agents on the shelf