china-qijizhifeng/agentic-harness-engineering

What it solves

It addresses the difficulty of manually optimizing the "harness" (the surrounding infrastructure) of coding agents. Instead of fine-tuning the base LLM, which is computationally expensive, AHE automatically evolves the system prompts, tool descriptions, tool implementations, middleware, skills, sub-agents, and long-term memory to improve the agent's performance on coding tasks.

How it works

AHE uses an iterative evaluate → analyze → improve loop based on three observability layers:

1. Component Observability: Uses NexAU to decompose the harness into seven git-tracked, auditable components.
2. Experience Observability: An Agent Debugger distills massive raw execution traces into sourced reports that identify root causes of failure.
3. Decision Observability: An Evolve Agent proposes evidence-backed edits to the harness, predicts their impact, and uses subsequent evaluation results to falsify or confirm those predictions.

Every rollout runs in an isolated E2B sandbox to ensure safety and reproducibility.

Who it’s for

AI engineers and researchers developing coding agents who want to systematically improve agent performance without retraining the underlying base model.

Highlights

• Fixed Base Model: Optimizes the environment and tools rather than the model weights.
• Evidence-Driven: Every change must be backed by failure evidence and a predicted impact.
• Cross-Model Transfer: Evolved harnesses can be transferred to different base models without needing re-evolution.
• High Performance: Demonstrated significant pass-rate increases on Terminal-Bench 2.0.