I build reliability infrastructure for AI agent systems and write about what I find. Most agent pipelines fail the same way: retry logic that treats every error identically, no budget on retries, no floor on Retry-After, no explicit intent before irreversible actions fire. I've spent two years building tooling for that failure surface — Pitstop Scan, Guard, and intent-gate. If your agents are silently degrading in production, this is where I think out loud about why.