GPT-OSS 20B Ran My Personal Agent for a Week: 312 Tasks, 97.4% First-Shot Tool Calls, Zero Frontier APIs
rasheed106 · reddit · 2026-08-25
A developer ran a bold experiment: taking an open-source assistant harness, stripping out every cloud API call, and making GPT-OSS 20B responsible for the entire agent loop, running 7 days on an M5 MacBook Pro from a compiled 12GB binary.
Results:
- 312 real tasks, 1,847 tool calls
- 93.2% completed without human takeover
- 97.4% first-attempt schema-valid tool calls
- 71 multi-step workflows, 4.1% retry rate
He started the week trying to find where GPT-OSS 20B would fail — and ended it cancelling his Perplexity Computer subscription.
More from coding & agent
- Dev pushes back on "AI coding means insecure apps": humans always shipped bugs — dreamwieber · 2026-10-15
- Eight non-negotiable principles for building AI that actually learns from experience — tokenbender · 2026-10-15
- Animated teardown of XBOW's Chrome full-chain exploit that won Google's first $250K VRP bounty — moyix · 2026-10-15
- pstack's durable core: 10-20 line skills that distill agent best practices into principles — HankYeomans · 2026-10-15
- 'Children of Elon' forum lets AI agents post and teach each other while humans watch — AiJohnAllen · 2026-10-15
- VoidZero open-sources Void: SDK, CLI and platform to deploy Vite apps on Cloudflare — cnakazawa · 2026-10-15