Talk: how to RL-train an agent running inside a harness you didn't write
SergioPaniego · x · 2026-10-08
Lisbon AI published Sergio Paniego's talk on training agents that run inside a harness someone else wrote. The core problem: most agents doing real work were built by others, and standard RL assumes you control every step when the harness already does. The talk covers how to train under those constraints. Video on YouTube.
More from coding & agent
- Goose: a heap-free systems language claiming to beat C++ and Rust on speed and memory — pbaylies · 2026-10-08
- Gefei launches SEO Agent: one prompt runs keyword difficulty, domain valuation and a full SEO toolbox — gefei55 · 2026-10-08
- AI agent + 5.3 hours of driving data reveal the optimal coffee lid direction: two o'clock — mariyaivasileva · 2026-10-08
- Claude Code mods go programmable; Agent Guard puts coding agents in microVMs — Arindam_1729 · 2026-10-08
- Probing 16 subreddits with an agent account: 7 were already banned, and a 200 submit doesn't mean the post survives — lulzxdxdxd · 2026-10-08
- 4 models, one two-file bug: 3/4 passed, 10x cost spread, and the cheapest run was the failure — lulzxdxdxd · 2026-10-08