What was the first production failure mode that broke your tool-calling AI agent?
When moving tool-calling AI agents from a local demo into production, the failures that actually page you are rarely "the model wasn't smart enough." Across recent production systems, three failure modes keep showing up
When moving tool-calling AI agents from a local demo into production, the failures that actually page you are rarely "the model wasn't smart enough."
Across recent production systems, three failure modes keep showing up:
- Timeout-after-commit on write tools: The HTTP request times out after the downstream API already created the ticket, sent the email, or wrote the rowβand a naive retry duplicates the side effect.
- Silent retry loops: An agent hits an empty or partial result and calls the exact same tool with identical arguments until it burns the token budget.
- Context compaction drift: Summarizing old summaries drops negative constraints ("do not touch table X"), causing the agent to re-propose rejected paths 50 turns later.
Question for engineers running agents in production: Which failure mode hit your team first, and what deterministic guardrail (idempotency keys, state machines, external state ledgers, or hard circuit breakers) actually stopped it?
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.