The model weights are a red herring. While the industry obsesses over parameter counts and benchmark leaderboards, a new data point from Terminal-Bench 2.0 reveals that the execution harness—the layer between the model API and the terminal—dictates actual performance more than the model itself. Claude Opus, a top-tier reasoning engine, delivered a 93% success rate inside Cursor but plummeted to 77% inside Claude Code. Same weights. Same training data. 16-percentage-point swing. The only variable was the wrapper.

The Wrapper Is the Agent

For those of us building agent stacks, this distinction is critical. The 'harness' includes the system prompt injection logic, the tool definition schemas, the context window management, and the error-handling loops that feed outputs back to the model. Cursor’s implementation of the Claude Opus API clearly optimizes for these factors better than Anthropic’s own Claude Code wrapper in this specific terminal-benchmarking context. This isn't just about prompt engineering; it's about the architectural integrity of the agentic loop.

Benchmarking the Wrong Layer

We need to stop evaluating AI coding assistants solely on raw model capability. If a benchmark like Terminal-Bench 2.0 shows a 16% variance based on the client implementation, then 'model choice' is a secondary optimization problem. The primary engineering challenge is the harness. It determines how effectively the model can read files, execute commands, parse errors, and iterate. A mediocre model with a robust, well-tuned harness will often outperform a frontier model trapped in a poorly designed wrapper.

Key Takeaways

  • Execution harness quality is the new bottleneck for AI coding agents, not model intelligence.
  • Claude Opus showed a 16% performance drop moving from Cursor to Claude Code on Terminal-Bench 2.0.
  • Tool schema design, context management, and error handling within the wrapper define real-world agent capability.
  • Developers should audit their agent harnesses before assuming model upgrades will solve workflow failures.

The Bottom Line

Stop praying for better models. Start fixing your harness. The wrapper is where the agent lives, and right now, most of ours are poorly optimized cages for very smart animals.