Lesson 22 / 25

Failure Handling and Stop Conditions

Add retries, timeouts, step limits and human checkpoints so agents fail safely.

By · 15 years in web development

Plan for things going wrong

Agents loop, retry the same failing step, or wander. Defend with a maximum number of steps, a time limit, a cost cap, and checkpoints where a human approves risky actions such as deleting data or deploying. When a limit is hit, stop and report clearly instead of continuing.

A guarded agent loop

step runs one model-and-tool turn and is_done checks the result. The loop cannot run forever or spend without limit.

MAX_STEPS, MAX_COST = 15, 2.00   # dollars
spent = 0.0
for i in range(MAX_STEPS):
    result, cost = step(state)
    spent += cost
    if is_done(result):
        break
    if spent > MAX_COST:
        raise RuntimeError("Cost cap reached; stopping for review")
else:
    raise RuntimeError("Step limit reached without finishing")

Make failures visible

Log every tool call and decision. When something goes wrong at 3 a.m., a readable trace is the difference between a ten-minute fix and a day of guessing.

Quick check: Which is a good stop condition for an autonomous loop?

  • Run until the model says it is finished, no matter what
  • A maximum number of steps plus a cost cap
  • Never stop
  • Stop at random
Answer

A maximum number of steps plus a cost cap — Hard limits guarantee the loop ends even when the model never decides it is done.