Harness Engineering for Self-Improvement
Lilian Weng surveys the emerging research field of harness engineering for recursive self-improvement (RSI). A harness is the system surrounding a base model that orchestrates execution — how the model thinks, plans, calls tools, manages context, stores artifacts, and evaluates results. The core argument: near-term RSI won't come from models rewriting their own weights, but from harness engineering evolving as a meta-methodology — improving the machinery for getting better answers, not just improving individual answers.
Harness Design Patterns
Three patterns recur across the literature: (1) workflow automation — goal-oriented loops of plan, execute, observe, improve, repeat; (2) filesystem as persistent memory — durable state in files rather than everything in context, since artifacts grow much longer than context windows; (3) sub-agents and backend jobs — explicit, inspectable parallelism where subagent outputs live as files/logs, not transient chat context. The coding agent case study shows tool convergence across Claude Code, Codex, OpenCode, and Cursor: filesystem tools, shell execution, git, MCP/skills, web search, artifacts, backend processes, and agent delegation.
Context Engineering: From ACE to Meta-Harness
Agentic Context Engineering (ACE; Zhang et al. 2025) treats context as an evolving playbook of bullet points with generators, reflectors, and curators — but the update rules remain handcrafted. Meta Context Engineering (MCE; Ye et al. 2026) separates mechanism from artifact, running bi-level optimization: a meta-agent performs crossover over prior skills, while a base-level context engineer learns context functions from rollout feedback. Meta-Harness (Lee et al. 2026) moves another level deeper — the optimized object is the code that determines what information should be stored, retrieved, and presented.
Workflow Design as Search
Rather than handcrafting workflows, ADAS (Hu et al. 2025) formulates agent design as meta-agent search over code, and AFlow (Zhang et al. 2025) represents workflows as graphs optimized via Monte Carlo Tree Search. Both outperform manually designed workflows.
Self-Improving Harnesses
Self-Taught Optimizer (STOP; Zelikman et al. 2023) showed recursive improver improvement with a cautionary result: it worked with GPT-4 but degraded with weaker models — recursive structure alone isn't enough. Self-Harness (Zhang et al. 2026) runs a propose-evaluate-accept loop: mine failure patterns, propose bounded edits, validate with regression tests. Agentic Harness Engineering (AHE; Lin et al. 2026) focuses on observability — every failure must map to a specific harness component, every edit must be a falsifiable claim. Ahe achieved better than human-designed harnesses on Terminal-Bench-2.
Evolutionary Search
Darwin Gödel Machine (Zhang et al. 2025) explicitly targets harness-code evolution: agents propose improvements to their own harness codebase, and only high-performing children survive. With Claude 3.5 Sonnet, DGM-discovered agents went from 20% to 50% on SWE-bench Verified. Variants include AlphaEvolve, ShinkaEvolve (sample-efficient exploration), ThetaEvolve, and DemoEvolve.
Joint Optimization with Model Weights
SIA (Hebbar et al. 2026) combines harness improvement with model-parameter updates in the same loop, though Weng flags confounding experimental choices. Continual Harness (Karten et al. 2026) experiments with co-learning a policy model alongside harness updating.
Seven Future Challenges
- Weak and fuzzy evaluators (research taste, novelty)
- Context and memory lifecycle (context engineering as core intelligence)
- Negative results (publication bias → LLMs bad at abandoning hypotheses)
- Diversity collapse (evolutionary loops exploit known high-reward patterns)
- Reward hacking (unit tests, judge models, benchmark artifacts)
- Long-term success (maintainability, backwards compatibility, migration cost)
- The role of humans (oversight at the right abstraction level, not removal)