Meta FAIR's Self-Direction turns agent failures into reusable LoRA skills, +13.8pp on Terminal-Bench
karen_ullrich · x · 2026-10-27
Researchers from Meta FAIR and Princeton introduce Self-Direction: when an agent fails, it turns self-reflection into a small LoRA "skill" that fixes the retry, joins a growing library, and is reused on related tasks later—while the base model stays frozen. This addresses the common problem of agents fixing a mistake by retrying, then forgetting the lesson and repeating it next session.
Key results (base model Qwen3.6-27B, mean of three full runs):
- Terminal-Bench 2.1 task success: 65.5%, +13.8pp vs base
- Claw-Eval success: 73.3%, +8.9pp
- GAIA: 2.1 rollouts per task, 65% fewer than pass@6 at the same 55.8% accuracy
- 97.1% of later failures reused a stored skill (GAIA stream, 3-adapter library, no new training)
Paper and implementation coming soon; project page at self-direction.github.io. Core motivation: modern agents' recovery is ephemeral—extra compute vanishes with the episode, making long workflows costly and inconsistent.
More from coding & agent
- Kolab Benchmark Uses Pro Screen Recordings to Show AI Agents Overestimate Collaboration Quality — minnesotanlp · 2026-10-27
- Dev argues Claude burns far more tokens than GPT models for the same coding work — haider1 · 2026-10-27
- Stop building agents that can't change models: stable contracts, external state, real evals — rseroter · 2026-10-27
- Cheap models can cost more in agent harnesses than frontier ones, researcher warns — vasuman · 2026-10-27
- Kanzo: A 5k-Line Rust Kanban Binary With MCP for Claude Code and Codex — allmudi · 2026-10-27
- 'A human approves everything' is a trap: approval fatigue and how to actually design safe agents — Efistoffeles · 2026-10-27