API dies mid-task: only Atomic Agent with local fallback aced the test, Codex just hung
Ok-Brush8377 · reddit · 2026-11-06
A Reddit user stress-tested 6 coding agents by cutting the cloud API 45 seconds into a 5-part Python task with 50 hidden tests (3 runs each, local Qwen 3.8 27B Q4 on llama.cpp as fallback).
Results
- Without fallback, everyone failed: Codex hung on "Reconnecting..." until the cap (0/50×3); Claude Code gave up after 2.5 min — its fallbackModel doesn't trigger on connection errors; OpenCode, OpenClaw, and Hermes also bailed, all scoring 0.
- With fallback configured: Atomic Agent 0.6.13 (fallback on by default) switched in 2s and went 50/50/50; OpenClaw (80s to switch) went 50/33/33; Hermes (seconds) went 29/24/27.
Key findings
- The same 27B from scratch only scored 13/27/29, but with 45s of cloud planning before the cut it went 50/50 three times — the plan was what the 27B was missing.
- Codex is the only agent that never surfaces the failure; unattended, a hang is worse than a crash.
- The post includes concrete fallback configs for Hermes (fallbackproviders YAML) and OpenClaw (JSON), plus docs links; next round will test 429 rate limits.
More from coding & agent
- Indie dev: the key to better agent workflows is ceding more 'safety margin' to AI — yihui_indie · 2026-11-06
- Managing all your AI agents in one place: a screenshot tour of approval-gated computer use — RileyRalmuto · 2026-11-06
- Managing 30,000 Notes with Local Hybrid Search: qmd, MCP Servers and AI Assistant — dSebastien · 2026-11-06
- Skillware framework claims 97.3% cost cut as 3B-8B local models beat frontier LLMs — RossPeili · 2026-11-06
- A notebook app turns Claude into a tutor on reMarkable and iPad — tokentrillionaire · 2026-11-06
- Measured: in-process BM25 memory costs ~4.6KB RAM per entry, linear growth — Neither-Witness-6010 · 2026-11-06