FREE LLMs & agents LESSON · LLMs & agents
Evaluate the whole agent
Tasks, adversaries, budgets, and safe stopping
An agent succeeds only when the external state ends correctly.
Token-level quality cannot establish task success. Evaluation needs representative starting states, allowed tools, observable terminal conditions, side-effect checks, cost and latency budgets, and adversarial cases. Record the complete trajectory so failures can be assigned to perception, planning, policy, tool execution, verification, or stopping.
Judge the resulting state and path, not whether the transcript sounded intelligent.
A loop without a stopping proof spends risk as well as tokens.
Repeated planning can duplicate effects, drift from the objective, exhaust budgets, or keep acting after success. Bound turns and cost, make effects idempotent, detect unchanged state, require approval at irreversible transitions, and define safe failure terminals. A verifier must inspect actual postconditions rather than accept the model’s claim that work is complete.
Every loop needs progress evidence and an explicit reason to continue.