Guided · intermediate · 35 min · Premium

Give the agent a stop rule

Calder's support agent ran 47 turns on one case overnight, calling the same tool against a service in maintenance. Read the trace, bound the loop, handle the tool state the agent never handled, and replay the case to a clean handoff.

Brief

Customer: Calder Equipment (synthetic customer from the Production AI Agents course; every trace, log and config on this box is made up for the exercise)
Situation: The nightly agent run on calder-agent-01 processed eleven support cases. Ten finished in under five turns. Case C09 ran 47 turns before the harness aborted it, spent more tokens than the other ten together, and left 47 identical still checking notes in the operator queue. The replacement service was in its Tuesday maintenance window at the time.
Your role: You are the FDE who owns the agent loop. You may read the traces and budget log, run the agent's CLI (agentctl) to replay recorded cases with recorded model decisions, and edit agent/agent.yaml. Model outputs are recorded, so a replay is deterministic and costs nothing.
Goal: Name the loop shape from the trace, bound the loop with a turn budget, make the agent handle the unavailable tool state deliberately, state what a budget stop must do, and replay C09 to a recorded handoff in a handful of turns.

Objectives

1.Name the tool whose repeated call produced no new observation after turn 3 (you commit to an answer)
2.Bound the loop with a turn budget that a healthy case would never hit
3.Handle the unavailable tool state by stopping with a recorded reason and an operator handoff
4.State what a budget stop must produce (you commit to an answer)
5.Replay C09 to a recorded handoff in fewer than five turns, and let the live runs settle to handoffs

You will practise

→Read an agent trace and name the repeated observation that never changes the plan
→Set a turn budget and say what hitting it must produce
→Handle every tool availability state explicitly, not just success and error
→Prove a loop change by replaying the recorded case that exposed it

The mechanism, explained

The ReAct loop
the thought, action, observation cycle that C09 got stuck in
Long-horizon agent loops: stop conditions and evaluation
the interview question on budgets, stop rules and what to record when you stop
Course lesson: release with a working stop
a stop is a feature with an owner and a test, not a constant
Agent frameworks and their failure modes
the runaway loop is the first failure mode every framework guide lists, and why frameworks do not fix it for you
Agent loop checks with a decision model
where a small deterministic check belongs in the loop, before the model is asked again

Do first

The note that says ignore policy

Mission workspace

This mission is part of premium, with every challenge and incident in the lab. The brief above is the whole problem; the box, the hints and the debrief are what you unlock.

Get full Premium access · ₹2,000 / $25

Every answer, concept and course, plus Premium PDF guides, companion files, the full practice-test bank and work-sample downloads. Referral Premium excludes guide PDFs and their companion files.

6 months · One payment · No auto-renewal

Study alongside free video lessons.

← all missions next: Structured output, or nothing →