Failure Mode Weekly — Page 3

Agents pass staging tests but fail on real queries—here's how to catch the gap.

Agents fail to select the right tool as available tool sets scale beyond dozens.

Agent systems amplify retry costs because LLMs resubmit full context and spawn shadow retry loops.

Silent failures poison multi-step agents long before anyone notices the wrong answer.

Agents silently adapt to broken schemas where typed clients would crash.

Corrupted examples in agent prompts silently reshape model behavior across every downstream call.

Treats prompts like deployed code, not text files, with versioning and rollback built in.

Treating prompt injection as a one-time security fix leaves production agents exposed.

Models infer role from token patterns, not from API labels that developers assign.

The harness, not the model, causes agent paralysis and refusal.

Mixing delimiter styles in prompts creates vulnerabilities that modern attacks exploit at scale.