Agents break in production for boring backend reasons, not model reasons
DayKnown1050 · reddit · 2026-10-21
A backend/distributed-systems engineer argues that when agents fail in production it's rarely the model — it's the plumbing. Six fixes to prioritize before touching prompts:
- Make tool calls idempotent: agents retry; pass an idempotency key on every state-changing action so a timeout doesn't create two invoices.
- Separate decide from do: the agent proposes, a normal worker executes; anything moving money or deleting data goes through approval.
- Cap everything: max steps, tokens, and cost per run — one bad loop at 3am can burn your monthly budget.
- Save state after every step: resume at step 7 instead of redoing 1–6.
- Log every tool call with inputs and outputs: the only way to debug "the agent did something weird."
- Turn every failure into an eval: replaying real failed runs beats prompt tweaking.
More from coding & agent
- Neo4j GraphAcademy launches AI proxy so workshop learners skip OpenAI billing setup — JeremyCMorgan · 2026-10-21
- Three weekend Claude Code projects: personal benchmarks, mini-SaaS, AI asset library — EXM7777 · 2026-10-21
- vLLM gains sleep mode on Apple Silicon: idle model servers now return KV cache memory — abhijithneil · 2026-10-21
- Open-source Diagram Design turns prompts and codebases into editorial-grade technical diagrams — bibryam · 2026-10-21
- Agent web-data startup raises $9M led by General Catalyst, 800+ customers — aftahi_ai · 2026-10-21
- AWS shows how to build conversational supply chain agents with Bedrock AgentCore and MCP — AWS ML Blog · 2026-10-21