Agent harness engineering: lessons from 57 papers — same model, 6x performance gap
Marmelab · reddit · 2026-10-30
Synthesizing 57 papers plus hands-on experience, the author distills agent harness best practices: mine past sessions for rules, delete unproven rules, keep instruction files 100 lines (OpenAI's giant-file failure), use fewer sharper tools (Vercel cut 80% of tools, success 80%→100%, latency 724s→141s), enforce rules with scripts (only 4-16% of prose security rules actually work), and regularly prune stale rules — OpenAI even runs a 'doc-gardening' agent. Same model can vary up to 6x in performance by harness design (U. Toronto).
More from coding & agent
- WhatsApp outreach agent: one promo per person per 14 days, counted across all DM lanes — petrusenko_max · 2026-10-30
- Nova v1 'Flash' launches as an AI engineering teammate aiming to replace the IDE — Gauri_the_great · 2026-10-30
- Can You Pass This 5-Question PyTorch Interview Speedrun? eval, zero_grad, detach and More — cneuralnetwork · 2026-10-30
- Full AI Short Film Workflow: Character/Props/Location Sheets + Seedance Reference Mode — socialwithaayan · 2026-10-30
- David Khourshid's top prompting trick: append 'Ask Fable what it thinks' — DavidKPiano · 2026-10-30
- Harness launches E2E-encrypted orchestration for AI coding agents across your machines — dee_hw · 2026-10-30