AI Build Lab / Learn · Agent Workforce · Week 3 · Module 03 · Details

More detail: Build the team with Gigawatt

Optional reading for this lesson. Everything you must do or decide is in the lesson.

Week 3 · Module 03 · Lesson 3 · Details

This page goes with the lesson "Build the team with Gigawatt" (course/workforce/week3/lessons/anw-w3-l07.md). You do not need it to finish the lesson. Read a section when the lesson links to it.

The eight phases

Building mode is the Operator (Gigawatt) running the Fisk Method's eight phases out loud, with the phase block at the top of every building turn.

  1. Prime: checks that your approval is recorded and that no design file changed since your yes.
  2. Align: asks only what the design left open: sources, tools and edge cases.
  3. Research: one research file for each new agent; a web fact only if needed.
  4. Generate: builders stage each new agent.
  5. Evaluate: the thin slice, then the independent check and the rehearsals.
  6. Iterate: version 1.1 from the scorecard, kept only if it scores higher.
  7. Document: each new agent's notes and deployment notes.
  8. Deploy: the install, on your yes or your pre-authorization, then the proof in a new thread.

Coming back to a build you stopped

Gigawatt saves the next unfinished step in runs/<job>/build-log.md. In a new thread, give it your pointer line again. It checks your approval first, then picks up from the build log. Nothing is installed until your yes to a numbered install, so stopping partway leaves your apps unchanged.

Tool names, and when tools are granted

Gigawatt checks each tool's exact name on your app's connector page. If it cannot see the page, it asks you to read the names off it, word for word. A guessed name can quietly point an agent at the wrong tool, and your yes cannot fix that.

In Align, tools are only chosen. They are granted at the end, with the install. You connect every tool yourself; Gigawatt never signs you in or asks for a key.

Questions that would change the design

The design is the Architect's (Archie's). A question like "should the Trend Picker also write the intro?" changes it. So Gigawatt writes the question to a file in your design folder, and a fresh copy of Archie answers in a file. You never carry messages between threads.

A web fact goes the same way: one question, through a fresh copy of Archie, to the External Researcher (Scratch). Nothing private goes into a search.

How the builders work

Gigawatt starts one builder per new agent, on faster everyday models. Builders write only into the staging folder, runs/<job>/setup/. Each staged agent must pass the setup check with RESULT PASS. That line says the files are complete and shaped right, not that the agent does good work. A stronger checker judges that.

In Codex, builders may run one after another, because the app limits how many helpers run at once. Slower, not wrong.

The thin slice in full

The slice runs one real job, your test case 1, through your new lead and specialist, before the rest is built. Tools are not live yet, so it uses saved stand-in results for them. Its files are under runs/<job>/rehearsal/slice/: the request, the lead's plan, the evidence package and the stand-in tool results.

Gigawatt saves the tradeoff question, your answer and the result in build-log.md. Your words are kept apart from Gigawatt's notes on them.

When the build finds a gap in the design

Sometimes the build shows the design missed something real, like which agent puts the final piece together. Gigawatt writes a change note and questions for Archie in your design folder. A fresh copy of Archie answers in a file. It never talks to you.

  • "The design still holds." No new version, and your approval stands. This is a normal result. Gigawatt never invents a change.
  • A new version, for example prd-v3.md, with a short changelog at the top saying what changed and why. Gigawatt compares it with the version you approved on what the team may do: its tools, where it writes, what it sends or posts, which agents are borrowed, and the test cases. Nothing there changed: drift, noted in build-log.md and the install report, and the build goes on. Something there changed: the build waits, Gigawatt reads you the change, and you say yes to the new version in your own words, in Gigawatt's thread.

Gigawatt records which version your yes covers. Once you approve a new version, your old approval no longer works for the build.

The checker and its label

The checker starts on a stronger model, clean slate. It reads only the staged files, your approved design and a marking guide the builders never saw. It writes findings first, then its verdict, to runs/<job>/setup/checker-v<n>.md.

The label says how independent it was. Independent check: the same model family as the builders, clean slate. Second-model review: a different model family. Gigawatt records which one ran.

Rehearsals in full

Each new agent is tried on two test cases: case 1, the ordinary job, with the slice's real inputs, and case 5, the one the team must refuse. A stand-in helper runs the staged instructions with stand-in tool results and no live tools.

For each one, Gigawatt saves a brief and a report under runs/<job>/rehearsal/. The report sets the case's "Should" and "Fails if" lines from your design beside what the agent produced, and ends in PASS, CONCERNS or FAIL. It has a "Could not verify" section, and a section, "Differs from the native seat", that lists what the stand-in did not have: its tools, its skills, its notes, its model.

After the checks, Gigawatt writes version 1.1 from the scorecard and the findings, adding rather than cutting, and rechecks the changed files. Anything still marked FAIL comes to you as one decision before Deploy.

The install in full

The list. Gigawatt writes an install manifest and shows you every staged file, each skill, the tool grants you chose in Align and each model choice. If the Quality Reviewer (Kansa) reviews your team, a review skill for Kansa may be listed too.

Why a number. Every install list has a number, and Gigawatt ties your yes to it, so your yes covers exactly the files on that list. You say yes in your own words; Gigawatt writes down the number. A maybe ("not yet", "hold on") is not recorded.

The pre-authorized install. If you said the pre-authorization line at your design yes, Gigawatt does not ask. It reads three things from the files: every new agent's render check is PASS, no BLOCKING finding is open in the checker's file, and every rehearsal report's verdict is PASS. All three: it records the install on your pre-authorization, with the date you gave it, and installs. Any one missing: it names it and waits for your yes.

Your app's prompt. Writing into your agent folders makes your app ask once more: Claude Code before writing into .claude/, Codex before writing into its read-only agent folders. It is the mechanical gate for the yes you gave: one click.

The recheck. Before copying anything, Gigawatt checks your approval and the install list against the files on disk again. If any staged byte changed since your yes, it copies nothing, names the changed file, and offers a new list. If the design drifted in wording only, it goes on and quotes the drift note; if the drift changed what the team may do, it stops for a fresh yes.

The record. After the install, the Charter Steward proposes your team's record from the design and the registry, and your Chief of Staff files it with the install receipt, your yes as its owner_yes. The team check runs and its OK line goes into the receipt. No separate yes: the record only writes down what you just installed.

The two record checks. Drift: 0 findings means your rules and your agents' files agree, and nothing got a power your rules don't allow. Seat records: ... 0 mismatch(es) means every agent's tools match everywhere they are written: your list of agents, each agent's own files, and your roster page.

In Codex. Gigawatt may say the copy for Claude Code's @ list waits until you next ask it in Claude Code. That is expected.

Proving each agent

Each app reads its list of agents when a thread starts, so the install thread cannot reach your new agents. "Who are you?" only shows the app can find an agent. Proof is the agent doing its job: every new agent has a "prove I work" check in its design file, agent-<role>.md, with one input and the shape of the answer it should give.

An agent that does not answer, or does not do the job, stops the proof, and Gigawatt names what is missing.

Gigawatt's last proof step is one ordinary job through your Chief of Staff. Your Chief sends work to a team only from the team's record, which was filed with the install. A Chief that says the team is not on the record means the filing did not happen: ask Gigawatt, in its thread, for the install receipt and the team check line.