How AI Can Achieve Recursive Self‑Improvement: Lilian Weng Says Build a Robust Harness First

What Is Recursive Self‑Improvement?

The concept dates back to I. J. Good (1965) who described a super‑intelligent machine that can design better machines, and Yudkowsky (2008) later coined “recursive self‑improvement” (RSI) to denote a feedback loop where an AI uses its current intelligence to improve the mechanisms that generate intelligence.

In modern AI this loop can mean directly rewriting model weights or, more broadly, improving the training pipeline and deployment system that produce stronger successor models. Weng emphasizes the deployment layer—called the harness —as equally important as the raw model.

Harness Design Patterns

A harness wraps a foundation model, handling execution orchestration, planning, tool calls, context perception, artifact storage, and result evaluation. Successful coding agents such as Claude Code and Codex rely on a robust harness.

Pattern 1: Workflow Automation – Defines a goal‑directed loop (plan → execute → observe/test → improve → repeat). Karpathy’s autoresearch repository is a concise example.

Pattern 2: File System as Persistent Memory – Instead of stuffing the entire workflow into the prompt, persistent state is stored in files, leveraging the LLM’s ability to read/write via shell commands.

Pattern 3: Sub‑Agents and Backend Tasks – Harness can spawn parallel sub‑agents whose outputs are recorded as files or logs, making parallelism explicit and inspectable.

Case Study: Coding Agent Harness – Major coding agents (Claude Code, Codex, OpenCode, Cursor) converge on a stable interface that includes file‑system operations, shell execution, I/O tools, external context, web search, artifact generation, backend processes, and agent delegation commands.

Weng predicts a two‑step evolution: (1) harness engineering will become a “meta‑methodology” that improves answer‑finding mechanisms rather than answers themselves, and (2) mature harnesses will enable automatic research loops while preventing over‑engineering of the harness itself.

Harness Optimization

The optimization trajectory moves from prompt engineering → structured context → workflow → harness code → optimizer code, with stronger models tackling more complex, general objectives.

Context Engineering

Agentic Context Engineering (ACE) (Zhang et al. 2025) treats context as an evolving playbook, comprising a Generator (creates task trajectories), a Reflector (extracts insights from successes/failures), and a Curator (incrementally updates structured context without overwriting the whole prompt).

Meta Context Engineering (MCE) (Ye et al. 2026) separates the mechanism for managing context from the content of the context, allowing meta‑level skill evolution while the base level learns from execution feedback.

Meta‑Harness (Lee et al. 2026) pushes the idea further: the object of optimization is the code that decides what information to store, retrieve, and present to the model. Proposals for new harnesses are themselves generated by coding agents, yielding a Pareto‑front of harness candidates.

Workflow Design

The AI Scientist pipeline (Lu et al. 2026, Nature) automates idea generation, code writing, experiment execution, result analysis, manuscript drafting, and peer review. ScientistOne (Meng et al. 2026) enforces verifiable design constraints, while Autodata (Kulikov et al. 2026) manages challenger/solver hierarchies but is limited to weak‑solver fine‑tuning.

Other approaches treat workflow design as a search problem: ADAS (Hu et al. 2025) formalizes agent design as an optimization where a meta‑agent programs new agents; AFlow (Zhang et al. 2025) represents workflows as graphs and uses MCTS to improve them on QA, coding, and math tasks.

Self‑Improving Harnesses

Self‑Taught Optimizer (STOP) (Zelikman et al. 2023) receives an initial solution, utility function, and black‑box LLM, then returns an improved optimizer. Experiments show STOP discovers genetic algorithms, decomposition, multi‑armed prompt bandits, simulated annealing, and beam/tree search, but only improves downstream performance on strong models (GPT‑4) and degrades on weaker ones.

Self‑Harness (Zhang et al. 2026) introduces a propose‑evaluate‑accept loop where agents (1) cluster failures by validator‑detected patterns, (2) propose bounded harness edits, and (3) validate edits via held‑in and held‑out regression tests before merging. Weng warns that allowing programs to edit OS‑level code blurs abstraction boundaries, requiring external permission and safety layers.

Evolutionary Search

Evolutionary methods excel when the search space is huge and gradients are unavailable but solution evaluation is cheap. Examples include Promptbreeder (Fernando et al. 2023), GEPA (Agrawal et al. 2025), AlphaEvolve (Novikov et al. 2025) which marks code diffs with # EVOLVE-BLOCK-START and # EVOLVE-BLOCK-END, ThetaEvolve (Wang et al. 2025), ShinkaEvolve (Lange et al. 2025) that balances exploration/exploitation and rejects overly similar candidates, and the Darwin Gödel Machine (Zhang et al. 2025) that lets agents edit their own harness repositories, achieving 20 %→50 % improvement on SWE‑bench Verified and 14.2 %→30.7 % on Polyglot.

Joint Optimization with Model Weights

SIA (Hebbar et al. 2026) combines harness improvement and model‑parameter updates in a single loop involving a Meta‑Agent (proposes harness), a Task‑Specific Agent (executes tasks), and a Feedback‑Agent (decides whether to update harness or weights). Early results show the task‑specific agent is much weaker than the others, making conclusions tentative.

Future Challenges

Even with powerful harnesses, several bottlenecks remain:

Weak and vague evaluators – Many research claims lack fast, precise validators, limiting self‑improvement to domains with measurable metrics.

Context and memory lifecycle – Autonomous agents will accumulate memory; Weng argues context engineering must become a core intelligence component.

Negative results – Literature bias toward successes makes it hard for LLMs to learn when to abandon hypotheses.

Diversity collapse – Evolutionary and RL loops tend to converge on similar high‑reward solutions, reducing exploration.

Reward hacking – Agents may over‑fit to unit tests, judge models, or benchmark scores, so evaluators and permission controls should sit outside the evolutionary loop.

Long‑term sustainability – Coding agents can complete tasks but preserving a large, jointly maintained codebase over time remains an open problem.

Human role – Humans should move to supervisory positions, providing oversight at appropriate abstraction layers.

Useful Benchmarks (Appendix)

PaperBench : Reproduces 20 ICML 2024 Spotlight/Oral papers; best model (Claude 3.5 Sonnet) reaches ~21 % of PhD‑level performance.

CORE‑Bench : Evaluates reproducibility of 90 papers across CS, social sciences, and medicine (270 tasks).

ScienceAgentBench : 102 tasks from 44 peer‑reviewed publications covering math, chemistry, biology, geography.

RE‑Bench : Compares frontier AI agents with human experts; best AI agent is 4× faster in a 2‑hour budget but humans surpass agents over longer horizons.

MLE‑bench : 75 Kaggle ML competitions; o1‑preview + AIDE scaffold reaches at least bronze level in 16.9 % of contests.

KernelBench : 250 PyTorch tasks assessing LLM ability to write fast, correct GPU kernels.

Original article: https://lilianweng.github.io/posts/2026-07-04-harness/