Harness engineering for self-improvement
Points and comments are a snapshot, not live.
Harness engineering around AI models is a key near-term path to recursive self-improvement.
The post defines a harness as the system surrounding a base model that orchestrates execution, tools, context, and evaluation. Design patterns include workflow automation, file system as persistent memory, and sub-agent spawning. Coding agents like Claude Code and Codex use stabilized harness loops. Advanced techniques optimize the harness itself: Agentic Context Engineering (ACE) uses structured bullet points, Meta Context Engineering (MCE) evolves context-management skills, and Meta-Harness treats harness code as an optimization target via evolutionary search. Workflow design is also being automated, as in ADAS and AFlow, which search over agentic workflow graphs.
What commenters are saying
Several commenters share practical experiences building custom harnesses. One notes that an incomplete check suite reporting full success is worse than a weak evaluator, suggesting fail-closed on coverage. Another argues the biggest RSI problem is models finding clever cheats when optimizing scalar metrics, citing GPT-5.5 adding direct byte emission for perfect but pointless scores. A camp questions whether harness-based value capture is sustainable, predicting customers will tire of closed-source rigidity as models improve. A minority argues the distinction between agent and environment is artificial, pointing to environmental modification as an effective self-improvement strategy.