Lilian Weng’s latest Lil’Log survey reframes recursive self-improvement around the harness — “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.” Her near-term RSI prediction: it will not start with a model rewriting its own weights, but with the deployment layer becoming the optimization target, moving from prompts → structured context → workflow → harness code → optimizer code. The design patterns are strikingly close to how we actually run agents today: workflow loops (Karpathy’s autoresearch, the Codex agent loop), the file system as persistent memory (durable state in logs and files instead of context), and explicit, inspectable sub-agents and backend jobs. “Once harness design becomes an executable search space,” she writes, “a strong coding agent can exploit the same design space human engineers use.”
The survey maps the research stack: context engineering (ACE’s evolving bullet-point playbook, Meta-Harness optimizing the code that manages context), workflow design as search (ADAS, AFlow’s MCTS over workflow graphs), and self-improving harnesses — STOP (which notably degraded with weaker models: “the base model must be capable enough to improve the mechanism”), Self-Harness, and Agentic Harness Engineering, which beat human-designed harnesses on Terminal-Bench-2 by making every edit an observability-grounded, falsifiable claim with the verifier, model, and reasoning budget locked read-only. A striking finding from Lin et al.: harness-updating capability is flat from Qwen3.5-9B to Opus 4.6 — a 9B can write skills isomorphic to Opus — but harness-benefit (actually using the harness well) is what scales with intelligence. The open problems are the honest ones: weak evaluators, context/memory lifecycle, diversity collapse, reward hacking, and keeping humans “up the stack, not removed from the loop.”