What determines an AI agent’s performance is not only the large language model (LLM) itself. The harness, surrounding software that makes a model use tools, verify results and continue long-running tasks, is emerging as another key competitive advantage.

Gigazine, a Japanese media outlet, reported on Aug. 25 that AI software developer Earendil simplifies this as “agent = model + harness.” Models such as ChatGPT and Claude can understand questions and generate answers, but searching the web for up-to-date information, editing files, verifying calculation results and redoing work when there are errors requires a separate execution environment.

Earendil divided typical harness functions into four broad categories. The first is a system prompt that sets the model’s role and rules of behavior. The second is tools that perform external tasks such as web searches, code execution and writing emails. The third is an agent loop that repeats the process of “task → check result → decide next action.” The last is a conversion layer that aligns model-calling methods and data formats across different providers such as OpenAI and Anthropic.

Pi, an open-source harness developed by Earendil, also adopts this structure. It allows models from multiple AI companies to be switched and used in a single environment, and lets users tailor system prompts, tools and extensions to their way of working. Extensions shared by Pi users have exceeded 5,000.

An example showing the importance of harnesses is Nvidia’s AI agent system AVO. AVO combined Claude Opus5 with persistent memory, tool use and a supervision system to solve all 25 environments and 183 levels in the ARC-AGI-3 public set, scoring 100 points. Nvidia explained that memory retains prior results in long-running work, and the supervision system helps choose a different strategy after repeated failures.

Still, it should not be interpreted as the harness boosting performance more than threefold by directly comparing this with Opus5’s roughly 30% score. Nvidia also said the two evaluations had different conditions and the result was not a measurement that isolated the effect of the harness alone. Instead, it said the case shows it is difficult to assess the capability of a real agent system using only a model’s standalone benchmark.

Competition in AI agents does not end with securing more powerful models. Harness design is becoming another decisive battleground, determining what information to remember, what tools can be used, how to recover after failures and how to sustain long-running work.