AI workflows
How the workshop runs: agents that share one memory, write down what every test should show before it runs, score themselves against what really happened, and stop for a yes.
The record so far
- Sealed predictions
- 57 of 68 held-out predictions met, each written down and sealed before its test. 11 missed; 210 are still waiting for theirs.
- First prints
- 12 of 26 first prints met their acceptance criteria on the first try, and 7 more met them in part. Where the criteria were written before printing: 6 of 11.
- Shared memory
- 2,399 reads of the shared memory by agents across 20 sessions, 28 October to 11 November 2026, with 1,133 notes written.
One shared memory
Every agent session starts by reading a shared memory: a library of short notes, one rule to a note, filed by what it teaches rather than by the project that taught it. Agents write back what they learn, at every checkpoint, not at the end.
Unattended agents can only leave drafts in an inbox, and I promote or reject each one. Every note says how sure it is (hypothesis, sourced, tested, in use, validated), and that level rises only on evidence.
Predictions sealed before every test
Before a physical test, the agent writes down what it expects: the fit, the weight, the time, the way it will fail. The record is sealed, so it can't be edited after the fact.
After the test every prediction is scored met, partly met or missed, and every miss gets an explanation, which becomes a prediction of its own. Only held-out predictions count toward the headline; the ones used to set a number don't.
A routine that stops for a yes
The routine lives in skills: short written procedures any agent can run. A few of them:
- Say-backA picture of what will be made, and a yes, before any modelling.
- Render reviewThe proof sheet the yes is given on, in the real colours.
- Prove fitOne small test coupon for every fit that isn't proven yet.
- Print designThe route from idea to print files, with the predictions sealed on the way.
- Print reportScores the result against the seal and explains each miss.
- Brain logFiles the lesson where the next piece will find it.
- DecideA decision stays proposed until I approve it in my own words.
- CheckpointWrites the resume point, commits, and stops for a go.
Many agents, one plan
Big jobs are split across agents with narrow written briefs, and the riskiest steps get an adversarial check: independent agents try to refute a finding before it counts. Convergence Park, a math game about real animals, was built in a 48-hour run across 17 parallel workstreams.
This site was made the same way: a written plan approved first, nineteen candidate pieces vetted for licences and privacy, and independent reviews of privacy, accuracy and accessibility before anything went public.