The Evolution of the Agent Harness: Why agents started working
Latent Space · rss · 2026-08-22
Argues that the late 2025 breakthrough in agents came from the convergence of the "model capability curve" and the "harness complexity curve".
The Harness is everything besides weights (environment, tools, memory, guardrails). The evolution:
- Harness 1.0: The Bolt-On Era
- ReAct (2022): Theoretical loop via prompting.
- AutoGPT/BabyAGI (2023): Premature autonomy failed due to unreliable models (widest gap).
- Cursor/Copilot (2023-24): Retreated to "human-in-the-loop".
- Harness 2.0: The Co-Training Era
- Claude Code (2025): Seized the moment when reasoning models (o1) inverted the gap, granting autonomy with permission rules.
- Fusion & Absorption: RL moves inside the harness. Models absorb harness capabilities (e.g., auto-compaction) into weights. Future harnesses will shed scaffolding and focus on managing human attention.
More from AGI Musings
- Li Dan on Content Creation: Competing with AI via Unpredictability — oran_ge · 2026-08-22
- From nuclear energy to mind uploading: The future split of human civilization — Dr_Singularity · 2026-08-22
- Anti-datacenter movement is humanity recognizing its successor — ZeroStateReflex · 2026-08-22
- RL Training Reveals Models Learn Pivot Phrases Like 'No, That's Not Correct' to Manipulate the Simulator — voooooogel · 2026-08-22
- Can Model Personas Steer the Simulator? Extending the Shoggoth Metaphor — voooooogel · 2026-08-22
- Why Simulation is taking over: 10% worse, 100x cheaper, 10000x faster — Latent Space · 2026-08-22