Self-Direction adds 0.02% parameter adapters per lesson and lifts success 8.9-17.0 points on GAIA, Claw-Eval, Terminal-Bench 2.1

The method runs a single reflection pass on the failed trace, then trains one adapter equivalent to roughly 0.02 percent of base parameters while the original weights remain untouched. Adapters are indexed in a library and retrieved at inference time by task similarity. Experiments cover open-ended research, tool-use workflows, and terminal command sequences.

Across the three benchmarks the adapters raise end-to-end success between 8.9 and 17.0 percentage points. Repeated-mistake rate falls by nearly half; a small set of adapters transfers to dozens of unseen tasks without retraining. No full-parameter updates or context-window growth is required.

The approach decouples short-term retry from long-term policy change, sidestepping both context overflow and catastrophic forgetting. It directly addresses the documented failure mode in which in-context reflections disappear between independent sessions. Operational cost is limited to one lightweight training step per error and one retrieval lookup per query.

Next steps include scaling the adapter library to production agent fleets and measuring interference when hundreds of adapters are active simultaneously.