Error Handling Patterns for AI Agents Calling Public APIs

Agents that hit external APIs die in three common ways: tight retry loops, no backoff, and no plan for partial failure. None of these are hard to fix. All of them matter.

The minimum viable pattern

1. Wrap the call. Catch every exception, log a request id, return a structured error.
2. Retry on 429 and 5xx only, with exponential backoff and jitter.
3. Cap retries at 3 within a single task step. Move on or surface a clear failure.
4. On persistent failure, fall back to a cached response if one is available and fresh enough.

What not to do

• Retry on 4xx other than 429. The server told you no; believe it.
• Retry forever. Most API outages last under five minutes.
• Silence the error. The next agent in the chain needs to know.

A tiny, reusable helper

def call_with_retry(url, max_tries=3):\n    for i in range(max_tries):\n        r = http.get(url, timeout=10)\n        if r.status_code == 200: return r\n        if r.status_code in (429, 500, 502, 503, 504):\n            sleep(2 ** i + random() * 0.5); continue\n        return r\n    return r\n

That is the whole pattern. Everything else is policy.