My Brain Wiki

Harness Engineering for Self-Improvement

4分钟阅读

Harness Engineering for Self-Improvement

Author: Lilian Weng Date: July 4, 2026 Reading Time: 28 min Tags: language-model, agent, auto-research, self-improvement, prompting

Abstract

The concept of recursive self-improvement (RSI) dates back to I. J. Good (1965). This post focuses on research around harness engineering and how it contributes to RSI. A harness is the system surrounding a base model that orchestrates execution and decides how the model thinks and plans, calls tools and acts, perceives and manages context, stores artifacts, and evaluates results.

Harness Design Patterns

Compared with early agent frameworks (“agent = LLM + memory + tools + planning + action”), harness engineering additionally includes workflow design (e.g. loop engineering), evaluation, permission controls, and persistent state management.

Pattern 1: Workflow Automation

  • Goal-oriented loop: plan → execute → observe/test → improve → execute again
  • Example: Karpathy’s autoresearch repo
  • Model analyzes its own trajectories and failure cases via “agent runtime”

Pattern 2: File System as Persistent Memory

  • Harness should keep durable state in files, not carry everything in context
  • File read/write/edit via bash is a foundation skill for LLMs

Pattern 3: Sub-agent and Backend Jobs

  • Spawn multiple subagents in parallel, monitor backend jobs
  • Key: make parallelism explicit and inspectable (store as files/logs)

Case Study: Coding Agent Harness

  • Stabilized across Claude Code, Codex, OpenCode, Cursor
  • Tool groups: File system, Shell, IO (lsp, git), External context (MCP, Skills), Web search, Artifacts, Backend processes, Agent delegation

Harness Layer vs Core Intelligence?

  • Near-term RSI unlikely to start as model rewriting weights directly
  • Prediction: harness engineering evolves toward meta-methodology
  • Harness improvements may eventually be internalized into model behavior, but external interface remains

Harness Optimization

Progression: instruction prompts → structured context → workflow → harness code → optimizer code

Context Engineering

  • ACE (Agentic Context Engineering; Zhang et al. 2025): context as evolving playbook with Generator, Reflector, Curator; structured bullets merged deterministically
  • MCE (Meta Context Engineering; Ye et al. 2026): separates mechanism (how to manage context) from artifact content; bi-level optimization — inner optimizes context given skill, outer optimizes skill; uses free-form skills stored as files
  • Meta-Harness (Lee et al. 2026): optimized object is the code that determines what info to store/retrieve; proposer is itself a coding agent; Pareto frontier of harness candidates

Workflow Design

  • AI Scientist (Lu et al. 2026): pipeline for idea generation → code → experiments → paper → peer review
  • ScientistOne (Meng et al. 2026): verifiability as central design, Chain-of-Evidence checks
  • Autodata (Kulikov et al. 2026): challenger/weak solver/strong solver/verifier roles for synthetic data
  • ADAS (Hu et al. 2025): meta-agent search, agent design as optimization problem
  • AFlow (Zhang et al. 2025): workflow as graph, MCTS optimization

Self-Improving Harness

  • Code is universal language for defining programs/systems
  • STOP (Zelikman et al. 2023): recursive scaffolding improvement — improve the improver itself; discovered strategies like genetic algorithms, simulated annealing; caution: only works with capable base models
  • Self-Harness (Zhang et al. 2026): propose-evaluate-accept loop; learns model-specific harness instructions
  • Promptbreeder (Fernando et al. 2023): optimize prompts through mutation, mutation prompts also evolve
  • GEPA (Agrawal et al. 2025): reflection + evolutionary search
  • AlphaEvolve (Novikov et al. 2025): coding-agent evolutionary search, EVOLVE-BLOCK markers, meta-prompt co-evolution
  • ThetaEvolve (Wang et al. 2025): evolutionary search + RL + ICL
  • ShinkaEvolve (Lange et al. 2025): sample-efficient exploration, code-novelty rejection, meta-scratchpad
  • Darwin Gödel Machine (Zhang et al. 2025): evolution of editable harness-code; agent modifies its own harness
  • Hyperagents (Zhang et al. 2026): meta-agent controls how to modify task agents

Joint Optimization with Model Weights

  • SIA (Hebbar et al. 2026): Meta-Agent + Task-Specific Agent + Feedback-Agent; combines harness improvement and model-parameter updates

Future Challenges

  • AI Scientist can write papers but not necessarily do real science
  • Paper production ≠ scientific discovery
  • Reward hacking remains open
  • Permission control and security layers needed outside self-improvement loops
  • Compute efficiency of evolutionary methods

Key Benchmarks

  • PaperBench, RE-Bench, MLE-bench, ScienceAgentBench, CORE-Bench, KernelBench, TerminalBench, AIDE/Kaggle

关系图谱