Set goals, budgets, stop conditions, and escalation

Verifiedon 2026-07-302026.7.1-2

A goal needs a terminal state

A useful goal says what will be delivered, how it will be checked, and when the work stops. For the researcher: “By 10:00 UTC, return no more than eight official URLs from the approved source list, each with a one-sentence claim and retrieval time. Stop when the list is complete, the time limit is reached, or a source is unavailable.”

Do not define success as “the agent feels finished.” Define terminal outcomes: accepted, rejected with reason, timed out, or escalated. Each terminal outcome produces a durable record.

OpenClaw’s agent exec --json run envelope can expose status, usage when available, estimated costUsd when model pricing is known, assistant turns, and tool summary. Use those as operational inputs, not as proof that an output is correct.

Lab: write the control card

For one task, record:

• acceptance test and independent reviewer;
• deadline and wall-clock limit;
• cost ceiling or maximum turns/tool calls where cost is unavailable;
• prohibited actions and required approval boundary;
• stop conditions: missing receipt, repeated identical failure, exceeded budget, or unsafe input;
• escalation owner and required packet (artifact, error class, run ID, cost/usage, recommended next action).

Check: A different operator should be able to terminate the task from the card without asking its author what “enough” means.

Checkpoint: budget gate

A verifier has spent 80% of a $5 research budget and still has two unverified claims. The correct next step is not an automatic retry. Record partial evidence, stop the task at the threshold, and escalate the choice to increase budget, reduce scope, or hold the bulletin.

Learner artifact

Draw a state machine from queued → running → accepted/rejected/timed-out/escalated. Add red transitions for budget breach and unsafe-input detection. No state should point directly from running to published.