NOFire deleted most of its agents: single loop gave 1.5x diagnoses at 60% lower cost
tianyin_xu · x · 2026-11-05
NOFire's AI SRE team shares a counterintuitive engineering postmortem: their planner-worker-evaluator-writer multi-agent setup turned out to be the slowest, costliest and least accurate design they measured. Switching to a single agent loop diagnosed 1.5x more faults with 49% fewer tool calls, 60% lower cost per investigation, and 2.5x faster diagnosis, validated via daily evals on SREGym. CTO Panagiotis Moustafellos's takeaway: "Evals shouldn't just validate your architecture, they should try to break it." SREGym author tianyinxu amplified the post.
More from coding & agent
- Alibaba open-sources qz, a browser agent that runs inside your signed-in Chrome — thisdudelikesAI · 2026-11-05
- Rentcast MCP Server open-sourced for property data, valuations and rental estimates — modelcontextprotocol · 2026-11-05
- EIA API v2 MCP connector lets agents query US energy data via Streamable HTTP — modelcontextprotocol · 2026-11-05
- Uber open-sources ADR, its production agent security system with 300-task benchmark — thisguyknowsai · 2026-11-05
- Dev lets Codex log into his tax portal and file half a year of overdue returns, ¥30k+ paid — yihui_indie · 2026-11-05
- Tencent's Hy4 lands on Vercel AI Gateway at 50% off through end of 2026 — TencentHunyuan · 2026-11-05