Error Handling and Recovery · Agentic Systems

Agents fail in non-deterministic ways. Error handling has moved from "Try-Catch blocks" to Agentic Self-Correction and Stateful Rollbacks, with frameworks…

Agents fail in non-deterministic ways. Error handling has moved from "Try-Catch blocks" to Agentic Self-Correction and Stateful Rollbacks, with frameworks like LangGraph and Microsoft Agent Framework providing native checkpoint/resume primitives.

The Taxonomy of Agent Failures
Self-Correction Loops
Stateful Rollbacks (Checkpointing)
The "Stuck in a Loop" Fix
Graceful Degradation
Interview Questions
References

Taxonomy of Agent Failures

Hallucinated Tools: Calling a tool that doesn't exist.
Schema Violation: Passing the wrong arguments to a real tool.
Environment Error: Tool exists, but the external API is down.
Logical Stall: The agent performs the same failing action repeatedly (The ReAct Loop of Death).

Self-Correction Loops

Errors are now treated as Tokens of Information.

Pattern: When a tool fails, the error message is NOT just logged; it is fed back to the model as a prompt: "Action failed with error: X. Reflect on why this happened and provide an alternative strategy."
Reasoning Models (Claude Opus 4.7 extended thinking, GPT-5.5 reasoning, DeepSeek-R2): These models excel at this because they "internalize" the error during their hidden Chain-of-Thought, leading to a much higher one-shot recovery rate.

Stateful Rollbacks (Checkpointing)

For long-running agents, an error in Step 9 shouldn't crash the whole project.

Checkpoints: High-reliability systems (using LangGraph or similar) save the "State Snapshot" to a DB after every successful tool call.
The Rollback: If the agent enters a logical stall, the supervisor agent can Reset common-state to Step 5—the last "Safe" state—and force a different path.

The "Stuck in a Loop" Fix

Infinite loops are the #1 cost-sink in agentic systems.

Solution: Counter-Based Intervention.

If the same (Tool, Args) tuple is seen 3 times in one session, the orchestrator interrupts the model.
It injects a mandatory "Pivot Instruction": "You have tried searching for 'X' three times. This path is dead. You MUST try a different tool or admit you are stuck."

Graceful Degradation

If the high-reasoning agent (Claude Opus 4.7, GPT-5.5 reasoning) keeps failing, we fall back to:

Simplified Agent: A smaller model with fewer, more reliable tools.
RAG-only Mode: Disable actions and just provide a conceptual answer based on the knowledge base.
Human Escalation: (See the next chapter).

Interview Questions

Q: Why is traditional "Exception Handling" (Try/Catch) insufficient for Agentic Systems?

Strong answer: In traditional software, an exception is a "Stop" command. In an agentic system, the model is the "Driver." If the system just stops, the user task fails. We use Error Injection instead of Exception Handling. We catch the exception at the platform level and transform it into a Synthesized Observation for the model. This allows the model to "Reason" around the failure. A TRY/Catch only fixes the code; Error Injection allows the model to fix the Plan.

Q: How do you handle "Silent Failures" (Where the tool returns 200 OK but the data is wrong)?

Strong answer: Silent failures are the most dangerous. We implement Output Validation Agents. For critical steps, we don't just accept the tool output. We pipe the output to a "Verifier Agent" (often a smaller, faster model) whose only job is to check: "Does this tool output actually answer the query provided?" If the Verifier says "No," it triggers a self-correction loop as if it were a hard error.

References

LangGraph. "Persistence and Checkpointing" (2025)
Shinn et al. "Reflexion: Learning from Errors" (2024 update)
Microsoft. "Managing Hallucinations in Agentic Systems" (2025)

Next: Human-in-the-Loop Patterns