The last skill of the course: putting your two agents in roles — one builds, the other reviews — and seeing how far that idea scales. It scales further than you think: the course you're sitting in is built this way.
Run a real builder + reviewer cycle on your own capstone — Codex builds, Hermes reviews — then use the pattern in the clinic to close your project's remaining gaps before demo day.
You need: your feature-complete capstone from last night · both agents · your DONE WHEN scorecard.
Why two agents beat one: an agent reviewing its own work re-reads its own reasoning and nods along — the same blindness humans have. A different agent, with different training and no attachment, reads the actual files. The pattern is old (writers/editors, pilots/co-pilots); the price is new: your second reviewer is already on your subscription — a review spends some of your quota, not extra money.
Commission the review — like a job order
In Hermes, pointed at your my-tool folder:
You are the reviewer for this project. Codex built it; you did not — read it with fresh eyes.
Review against spec.md and AGENTS.md: 1) does every DONE WHEN check truly pass — test them; 2) what breaks with hostile input (empty fields, huge numbers, wrong types); 3) anything a first-time user would find confusing.
Output: a numbered list, most serious first, each item with WHERE and HOW TO REPRODUCE. Do not fix anything — reviewers report, builders fix.Step done when: you hold a numbered findings list. (An empty one means the review order was too soft — tighten and rerun.)
Send the findings back to the builder
In Codex, same folder:
A reviewer examined this project and reported the findings below. Address them one at a time: for each, tell me your fix, apply it, and I'll verify before you move to the next.
[paste Hermes's findings]Notice your own role change: you're not building and you're not reviewing — you're routing work between specialists and approving results. There's a word for that job: managing.
✦ Two agents, one review cycle, and your project just got better while you mostly read and nodded. Remember this feeling — it's the whole next section.
Step done when: the top findings are fixed and verified, one by one.
A true story, from inside this course. These lesson pages are produced by an agent team: one agent (Claude) designs the curriculum and writes specs and content; another (Hermes — yes, the same one on your laptop) executes bulk work like translations from written task files; a human reviews what lands on disk and routes corrections. The task files look like this:
# Task 025 — Russian translation of the attention-decay reader## Goal: translate readers/attention-decay.html → readers/ru/…## Rules: structure and code stay identical; only visible text changes…## Verify: tag parity EN vs RU; no English left in visible copy…## Report: RESULT: path, parity check, notable.
Read it again with today's eyes: it's a skill (Day 4) written as a job order, with guardrails and a report (Day 7), executed by a second agent and verified by a human (Day 1's pipeline, forever). Every course concept, stacked.
The one-person company, role by role. Click each card — this is theory today and someone's Tuesday next year.
That's the honest shape of the “agentic corporation” — not science fiction, just today's four patterns run permanently. The bottleneck isn't the agents; it's how well the director specifies and verifies. Which is precisely what you've trained for eleven days.
Different models have different characters, and a director casts them like actors. From our lab's daily experience:
The habit: notice HOW each model answers, not just THAT it answers. Code from the coder, words from the writer — pairing one of each is the most common two-agent team in the wild.
The clinic — close your gaps
Rest of the session: your capstone, with help in the room. Work the list in this order:
Step done when: your scorecard says all checks pass AND a stranger used your tool without asking you anything.
Tomorrow you present. Two minutes, three beats, no slides — your software speaks.
homework complete — mark the day done below
Tick when the homework above is finished — it syncs to your account.
Self-review re-reads its own reasoning and agrees with it. A different agent has no attachment and different blind spots — it tests the files, not the intentions. Same reason newspapers have editors who didn't write the article.
Routing and judgment: commission the review, hand findings to the builder, verify each fix, decide when it's done. Specify and verify — the pipeline's human half, now applied to a team instead of one agent.
The director's clarity. Agents multiply execution, not judgment: vague specs produce confident garbage at scale, and unverified results compound. The scarce skill is exactly what this course trained — precise orders and honest checking.