Model-Harness Fit

The execution harness is treated not as a mere shell but as the very contract the model was trained on. Running Codex CLI, Claude Code, and GitHub Copilot CLI on the same workstation produces different behavior for the same prompt, and this difference is bundled together as the model-harness-fit problem.

The difference is also clear in Terminal-Bench 2.0. As of April 30, 2026, Claude Opus 4.6 scored 79.8% when paired with ForgeCode and 75.3% when paired with Capy, and Cursor pushed its own agent from Top 30 to Top 5 simply by changing the execution harness.

The contract methods of the three execution harnesses also differ from one another.

• Codex operates on a typed asynchronous protocol that exchanges Submission and Event, and layers a separate JSON RPC layer on top for cross-process use.
• Claude Code is a direct conversation loop that processes an AssistantEvent stream, and its tool contracts such as edit_file differ from Codex's apply_patch.
• GitHub Copilot CLI uses a supervisor protocol where the host spawns a child process, controlling the agent via session.create and session.event.

Everything is interlinked—tool names, input schemas, citation tags, memory methods, system prompt structure, and even identity file conventions like CLAUDE.md and AGENTS.md. The conclusion is one: quality cannot be maintained by simply swapping out the model, and the model and execution harness must be matched as a pair.