How agents act

A Startup decides what to do next, acts within enforced authority, and learns from each result, with no human choosing its Tasks. These chapters explain the loop, its owners, its bounds, and what reaches the founder.

1. The operating loopEvery Startup and the Studio run one loop, wake, decide, act, review, with no human choosing the next Task.
2. What evals decideEvals grade every agent's work and gate every new version, and never change what waits for the founder.
3. Decision classesA decision class sorts every Verb by whether its effect can be undone and what is at stake, and the class sets how long an Action waits.
4. Delay windowsA delay window is the time a reviewed one-way Action waits before it runs, so someone can step in; if nobody does, it proceeds.
5. WatchersA watcher reads a namespace's pending feed and can pause, redirect or cancel any pending Action; every intervention is a Decision.
6. BrakesA brake stops one class of effects, such as sends, spend, publishes or price changes, across many Startups at once, until it expires or is lifted.
7. OversightOversight is the human review a piece of work requires, declared by what the work is, never by what an agent can do.
8. GuaranteesWhat holds for every Startup and every seat, under every operating model, enforced by the platform rather than by agent instructions.
9. Recovery and proofHow a Startup survives failed calls, restarts and bad Decisions, how reverts contain damage, and how the record proves the loop works.