Definition
Plain language
A system that automatically rewrites an AI agent's surrounding tools and prompts by watching it fail and learning from those failures.
As stated in the literature
A composable, adaptive, evolvable agent-harness foundry that decomposes the harness into typed processors and uses a meta-agent (AEGIS) to evolve it from execution traces, framing harness optimization as a Markov Decision Process with defenses against reward hacking, catastrophic forgetting, and under-exploration.
Also called: AEGIS
Why it matters: It improves an agent's reliability by fixing the scaffolding around it automatically, instead of requiring engineers to hand-tune prompts and tools after every failure.
For example, after an agent repeatedly fails to find a file, the system rewrites the agent's search tool and prompt so the next run succeeds.
Heard on the show
“It's called "HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundry.”Episode 147 — Agents Fail at the Body, Not the Brain: A Self-Rewriting Scaffold That Lifts a 9B Model 44 Points