When an Order Comes Back

Opening Scene

A dish sent back from the dining room doesn’t shut down the entire kitchen — a well-run service has a real, practiced process for handling it: understand what went wrong, decide whether to remake it, adjust for next time, and keep the rest of service running smoothly. An AI agent executing a multi-step plan needs this exact same practiced resilience for when an individual step fails, which, across enough steps and enough real-world tool calls, genuinely will happen.

In Plain English

Error handling in agentic systems covers how an agent responds when a step fails — a tool call errors out, a retrieved result is unusable, a sub-agent returns something malformed. Well-designed error handling includes retry logic (trying again, sometimes with an adjusted approach), fallback strategies (an alternative tool or method when the first choice fails), and graceful degradation (completing what’s possible and clearly flagging what isn’t, connecting directly to the constraint handling covered in Article 10).

The Old Way

Before agentic error handling matured, failures in multi-step automated processes were often handled far more brittlely:

• Traditional automated workflows, covered in Article 3, often halted entirely on an unexpected error, requiring manual intervention to resume.
• Early agentic experiments sometimes had no explicit error handling at all, simply failing silently or crashing the entire process on any single step’s failure.
• There wasn’t yet a well-established set of patterns — retry, fallback, graceful degradation — specifically designed for handling failures within an agentic loop.

Robust error handling patterns emerged as practitioners recognized that real-world tool calls and multi-step processes will inevitably encounter failures, and a resilient system needs to handle them productively rather than catastrophically.

What’s Changing (and Why AI Is the Reason)

1. Established retry and fallback patterns, connecting directly to the constraint handling covered in Article 10, have become standard practice in well-designed agentic architectures.
2. As agentic systems chain more steps together, connecting to the parallel execution covered in Article 11, robust error handling has become essential — a longer chain of steps has correspondingly more opportunities for something to fail along the way.
3. This connects directly to the human-in-the-loop escalation covered in Article 14, since a well-designed error handling strategy often includes knowing when to stop retrying automatically and escalate to a human instead.

The Metaphor, Fully Extended

The Kitchen	Error Handling Concept
A dish sent back, handled through a real, practiced process	A failed step, handled through defined retry and fallback logic
Deciding whether to remake a dish or adjust the approach	Deciding whether to retry a failed step or switch to a fallback method
Keeping the rest of service running smoothly despite one dish’s problem	Completing what’s possible while clearly flagging what failed
A kitchen’s practiced resilience to the inevitable occasional problem	An agent’s practiced resilience to the inevitable occasional step failure

Quick Recap

• Error handling covers how an agent responds when a step fails: retry logic, fallback strategies, and graceful degradation.
• Traditional automated workflows often halted entirely on error; agentic systems need more resilient, productive handling.
• This becomes essential as agentic systems chain more steps together, increasing the chance any single step fails.
• Well-designed error handling includes knowing when to escalate to a human rather than retrying indefinitely.

Where This Fits in the Series

Article 12 covered handling a failure resiliently. Article 13 covers something every one of these techniques depends on: an agent’s ability to actually remember what’s happened so far.