Agent harnesses are absorbing into model weights as capabilities converge

In late 2025, engineers observed a marked improvement in agentic systems, coinciding with the release of reasoning models and maturing harness frameworks. The improvement was not attributable to a single factor but to the convergence of two curves: model capability and harness sophistication. As models gained native tool-use and reasoning abilities, the historical gap between what harnesses demanded and what models could deliver began to close.

The concept of the "agent harness"—the environment, tools, context, and guardrails surrounding a model—has evolved through distinct eras. In the "Bolt-On Era," harnesses like ReAct (2022) and Toolformer (2023) relied on prompting or limited training to enable tool use, but models remained brittle and prone to error compounding when given autonomy. Attempts at full autonomy, such as AutoGPT and BabyAGI in early 2023, amplified these failures, leading to a retreat toward human-in-the-loop designs in 2023–2024, as seen in AI-powered IDEs like Cursor and Copilot.

By early 2025, reasoning models such as o1 inverted the dynamic: models began to outpace harness capabilities. Claude Code, released in February 2025, exemplified this shift by giving models direct access to tools (e.g., bash, file read/write) with permission rules replacing per-action human approval. Its rapid commercial success—reaching roughly $1B ARR within six months—was attributed to Anthropic’s timing: the crossover point where models could safely operate with autonomy had arrived.

The current era, "Co-Training," sees harness capabilities being absorbed into model weights. Reinforcement learning is now applied within harness environments, training models on real-world tasks across varied setups. OpenAI’s codex-1 (May 2025) exemplified this approach, and subsequent evaluations demonstrate measurable performance gains from harness design alone. On Harness-Bench, the same model scored between 52.4 and 76.2 across different harnesses—a 23.8-point spread—while on ARC-AGI-3, GPT-5.6 Sol’s score tripled from 13.3% to 38.3% with only retained reasoning and compaction improvements in the harness.

This absorption process is not merely theoretical. As models train within harness environments, they internalize capabilities such as auto-compaction of context and tool-use policies, reducing reliance on external scaffolding. The result is a blurring of the boundary between model and harness, with future agentic systems likely to be defined by how effectively models integrate harness-like functionality into their weights.