Research & Papers

Self-Improving AI Agents: Harness Experiments Reveal Systems Bottleneck

After 1,000+ harness experiments, continuous self-improvement hits a systems wall.

Deep Dive

Researcher Henry Pan details his experiments with self-improving AI agents that modify their own harness to solve terminal bench tasks. After experimenting for a couple of weeks, he concluded that continuous self-improvement is primarily an experiment-systems problem—the system needs a way to decide which improvements can safely compound. The findings reveal parallels to coding-agent customization (e.g., SKILLS.md) and highlight the challenge of building safe, compounding improvement loops. The writeup covers both successes and failures and is intended as a systems/research look, not a benchmark claim.

Key Points
  • Over 1,000 harness experiments tested whether an AI agent could self-improve its own execution wrapper for terminal tasks.
  • Continuous self-improvement is primarily a systems problem: safely determining which changes can compound is the main bottleneck.
  • Findings mirror challenges in coding-agent customization, where tools like SKILLS.md guide behavior without breaking existing functionality.

Why It Matters

Insights into safe, compounding self-improvement loops are critical for building autonomous agents that operate reliably over time.