How do you handle agent failures and implement error recovery?
LLM & AILevel: IntermediateSafety & Reliability
How do you handle agent failures and implement error recovery?
Hints
- Identify the stage: pretraining, adaptation, generation or evaluation.
- Name the inputs, constraints and a case where the approach fails.
Answer
Agents fail for predictable reasons: tool errors, invalid tool calls, context window overflow, infinite loops, budget exhaustion.
Error handling strategies:
try:
result = tool.execute(params)
except ToolError as e:
# Feed error back to agent
observation = f"Tool failed with error: {e}. Try an alternative approach."
agent.step(observation)
except MaxRetriesExceeded:
# Escalate to human or fallback
return escalate_to_human(task)
Loop detection: track the last N (action, observation) pairs. If the same action repeats with similar observations, the agent is stuck — interrupt and re-plan.
Budget enforcement: track token consumption and tool call count per task. If either exceeds budget, gracefully terminate and return partial results.
Source and licence: JustInCache & contributors — Awesome AI Interviews · Apache-2.0