Part 9 of 11 · 1 min read
Quick notes
- A goal for an agent is a contract with three parts: success criteria something other than the agent can check, limits on steps, cost and time, and stop and escalation rules.
- A run ends by budget, by a model's judgement, or by an external signal; production designs combine them.
- A claim of "done" is a hypothesis, not a result. An independent verifier cut false completions in the simulation from two fifths of tasks to under one percent.
- A budget is a limit, not a goal: it stops runaway cost but cannot tell a finished task from an unfinished one. A stall detector that triggers a replan and then an escalation is what raised correct outcomes most.
- Impossible tasks are why you need an escalation path, and the stall threshold is a trade-off between escalating possible tasks too early and never escalating impossible ones.