The harness
By the end of the term your harness includes, at minimum:
| Piece | What it is | Your version |
|---|---|---|
| Skill | A reusable, named behavior the agent can run | At least three, each doing one job in your team's automation |
| Hook | Fires automatically on an event, without being asked | At least two: one from Week 3, one your project needs |
| Rule | A standing constraint the agent must obey | At least three, including one that makes it argue with you |
| Memory | What the agent keeps across sessions | A memory design: what's kept, what's forgotten, why |
| Agent org | Sub-agents with specific jobs and a way to check each other | At least one sub-agent and one evaluator |
| Evaluation | How you know the thing worked | Twenty real inputs and a pass/fail check |
Each piece gets built on a Monday, demoed to a classmate before you leave, and written up in that week's log. Teammates' harnesses may share skills. Each student's harness must run the full automation on its own.
The harness is the venture. The rules encode your judgment. The hooks encode when your judgment is needed. The agent org encodes who you'd hire if you had money. Building it well is the entrepreneurship.
Monday by Monday
| Wk | You build | It's Satisfactory when |
|---|---|---|
| 1 | Agent setup | It runs on your laptop and you can say what it sees |
| 2 | First skill | It does something a one-off prompt doesn't, and you can say what |
| 3 | First hook | It fires on its own in front of a witness |
| 4 | Rules | You can show it arguing with you, and show the same input without the rule |
| 5 | Memory | The assumptions ledger lives in it; you can say what it forgets |
| 6 | Second hook | The one your project needs, firing on a real input |
| 7 | The harness | All pieces run your team's automation end to end on sample inputs. Harness v1 due Mon Oct 19. |
| 8–10 | Agent org, evaluators, hardening | A sub-agent with a job, an evaluator that catches a planted failure, twenty inputs passing |
| 11–12 | Deploy, hand off | Five days live; a colleague of the process owner can run it from your docs |
The Venture Log
Every Friday, through the course form, 500 to 800 words. Individual, even though the venture is shared: I want each person's account. Eleven entries.
Fields:
- Week number and date range
- Interviews and conversations this week (who, what you learned) and your team's running total
- Experiments and tests you ran
- Observations (what you saw, not what you think)
- Key decisions made, and the reason
- Assumptions tested (which one, how, result)
- Harness change this week (what you added or changed in your agent, and why)
- Touch-point status (which of the three run on real inputs)
- What's working · What's not working · One thing you'd do differently
- Book connection: one idea from the reading and how it touched your project
- AI audit trail: tools used, what you delegated, what you verified
Named entries (the form has no separate box for these — put each one in Key decisions, and why): 1 is your gift plan, 2 the gift response, 6 your pivot-persevere-abandon rationale after the alumni panel, 9 your post-competition synthesis, 10 your go-to-market and unit economics.
Log 1 has no assigned reading yet, so the book connection can be "nothing yet." Touch-point status will be "none" for the first few weeks; say so rather than skipping it. Your three inefficiencies go to Canvas, not the log.
Satisfactory means: evidence this week documented; decisions and pivots stated with reasons; reading-to-doing connection made; what's working and not working addressed; AI audit trail present; every field filled with concrete detail. "Nothing this week" is Needs Revision. "I tried X, it failed because Y, next week Z" is Satisfactory.
The two numbers
Two numbers follow your team from Week 3 to the end of the term: conversations with real people so far, and touch points running on real inputs. Every Venture Log reports both, and I'll ask for them out loud in class. Any team member may be asked, so everyone knows the whole story.
They're the fastest read on whether a project is alive. A team whose conversation count hasn't moved in two weeks has stopped learning, whatever else got built. The conversation floor itself is on the project page: two per team per week, Weeks 3 through 10, twenty minimum.