Self-Evolving Agent Harness

Agents can improve by rewriting their own prompts, memory and code while the model stays frozen.

77.58%

SpreadsheetBench accuracy, small model

A small model with an evolved harness beats a bigger model

How it works

Agents stop adapting
Deployed agents meet new tasks, but retraining weights is costly, opaque, and impossible through closed APIs.
Evolve the harness
Use run feedback to revise prompts, memory, workflows and code, then admit revisions only after fresh validation.
Gains with guardrails
Bounds cover generalization, forgetting and safe admission; cited cases show big gains but persistent errors spread.

From the paper