An essay traces the evolution of AI agent harnesses — the tools, context, memory, and guardrails surrounding model weights — from ReAct's prompting-only loop in 2022 through AutoGPT's premature autonomy, IDE-based human-in-the-loop retreat, and Claude Code's terminal-based autonomy breakthrough in 2025. It argues models are now absorbing harness capabilities directly into their weights (citing Harness-Bench's 23.8-point score spread from harness alone, and OpenAI tripling ARC-AGI-3 scores via harness changes), so engineers keep deleting scaffolding that gets absorbed. The piece predicts the harness's future role shifts from being an interface to the model toward being an 'attention-interface' governing when agents interrupt humans, predicting most agentic AI companies will ship a human attention policy surface analogous to AGENTS.md.
Table of contents
What a Harness Actually IsHarness 1.0: The Past, “The Bolt-On Era”Harness 2.0: The Present, “The Co-Training Era”Harness 3.0: The Future, “The Attention Era”The Attention-InterfaceQuestions this post answers
How much does the agent harness alone affect performance versus the underlying model?
Harness design alone can produce a large performance swing even with an identical model. Harness-Bench ran the same model across 106 tasks in different harnesses and measured scores ranging from 52.4 to 76.2, a 23.8-point spread caused entirely by harness differences rather than any change to the model itself. Engineers comparing agent harness designs can track findings like this on daily.dev.
What caused Claude Code to succeed where earlier autonomous agents like AutoGPT failed?
Claude Code succeeded because it launched at the point where model reliability finally caught up to the autonomy the harness demanded. Earlier tools like AutoGPT and BabyAGI granted full autonomy to models that were still brittle next-token predictors, causing compounding errors over multi-step tasks. Claude Code instead gave models bash and file access with permission rules once reasoning models like o1 made sustained autonomy viable, and it grew to roughly $1B ARR within six months. Developers evaluating autonomous coding agents can follow harness design shifts like this on daily.dev.
How much did retained reasoning and compaction improve GPT-5.6 Sol's ARC-AGI-3 score?
Retained reasoning and compaction alone tripled GPT-5.6 Sol's ARC-AGI-3 score from 13.3% to 38.3%, according to OpenAI's own reported results, without any underlying model change. This demonstrates that harness-level changes such as context retention strategies can produce performance gains comparable to or exceeding model upgrades. Teams tuning agent context and compaction strategies can track results like this on daily.dev.