Module 5.3 · Topic 4

Error Recovery and Resilience

Bottom Line Up Front: Agent failures fall into predictable patterns. Diagnosing root cause informs whether you need new instructions, corrected context, or a different approach. Design workflows with error tolerance so…

4.1 Common Agent Failure Modes

Agents fail in consistent patterns. Early recognition helps you diagnose root causes accurately.

Misunderstanding objectives: Agent understands parts but misses intent. Asked for strategic analysis, returns tactical procurement. Logic is sound given misinterpretation; problem is interpretation.
Hallucination: Agent states false facts confidently. "Revenue $2.3B" when actually $230M. Sounds authoritative but is wrong.
Incomplete execution: Agent completes 80% and stops. Specified 10 points; provides 7. Signals capability limit reached.
Format violation: Asked for table; returns prose. Requested 3 options; lists 8. Signals constraint parsing failure.
Reasoning error: Logic chain breaks. Correctly identifies trends but misapplies them. Reasoning is visible but flawed.

4.2 Diagnosing What Went Wrong

When agents fail, diagnose before demanding a redo. Understanding why failure occurred determines whether corrected input fixes it or a different approach is needed.

Re-read your original prompt. "Analyze customer segment" is vague; the agent might analyze profitability, demographics, or loyalty. Check facts: hallucination, outdated data, or missing spec? Look at scope: did the agent exceed it or undershoot? Different errors need different corrections.

4.3 Corrective Prompting and Re-Delegation

Corrective prompting is more efficient than starting over. Provide corrected information and regenerate.

Logic behind this approach:

Agents incorporate corrected input and adjust output while preserving valid prior work. Narrowing change to the failed element is faster than full restart.

Sample prompt:

Your pricing analysis covered only three of five competitors. I've attached the complete list with current data. Regenerate the pricing table with all five competitors, organized by price point (lowest to highest). Keep all other analysis unchanged.

What to expect in reply:

Agent adds missing two competitors, reorganizes by price, and preserves other analysis intact.

4.4 Designing Workflows That Tolerate Agent Errors

Design workflows that absorb agent errors. Use parallel paths, verification gates, and fallback options from the start.

Parallel execution: Run agent work in parallel with human backup for high-stakes work.
Verification gates: Evaluate deliverables before downstream delivery. Failed work goes back, not cascading.
Bounded autonomy: Low-stakes tasks run autonomous; high-stakes have approval gates.
Rollback: Stage output in sandbox. Revert if wrong without destroying dependent work.
Escalation: Agent flags cases where it hits limits for human review.

Topic 4 details

Topic overview

Topic details

4. Error Recovery and Resilience

Failure Patterns: Identifying common errors like hallucinations or incomplete execution helps quickly diagnose whether to fix the input or change the approach entirely.
Targeted Regeneration: Providing corrected input for just the failed element is far more efficient than forcing the agent to restart a task from scratch.
Workflow Resilience: Design processes with built-in rollback options, parallel human execution, and sandbox staging to absorb and isolate agent errors.

Key takeaway

Knowledge check

Error Recovery and Resilience: An agent confidently reports a company's revenue as "$2.3B" when the true figure is $230M. Which failure mode is this, why is diagnosing the root cause important before re-delegating, and what is usually the more efficient recovery than starting over?

Show answer

This is hallucination: stating a false fact confidently. Agent failures follow predictable patterns (misunderstood objectives, hallucination, incomplete execution, format violation, reasoning error), so you diagnose the root cause first because different errors need different corrections. Corrective prompting, through supplying the corrected data and regenerating only the failed element while preserving valid prior work, is usually more efficient than a full restart.