Harnesses

Part I is about the harness: the program around the model that runs the loop, holds the tools, decides what the agent may do and loads its context. Its seven chapters read the harnesses from their own documentation and source, each on a stated date and at a stated version, because these tools change month by month.

Start with what a coding-agent harness is. It names the four parts every coding agent has (the loop, the tools, the permission gate and the context), checks each against real harnesses, and ships a runnable version of about a hundred lines with a scripted stand-in where the model would be. The other chapters take the parts one at a time. The gate: permission models compared puts Claude Code, Codex CLI, Gemini CLI, the OpenHands SDK and Omnigent side by side, with a working settings file. The context: CLAUDE.md, AGENTS.md and .cursor/rules gives the load and merge rules with an audit script, and memory across sessions separates saved transcripts, compaction and the memory an agent writes for itself. The tools: an MCP server in Python builds a dependency-free stdio server and shows the JSON-RPC on the wire. Unattended use closes the part, with running an agent headless in CI and coding agent security: injection through tool results, MCP server trust, and secrets in configuration.

Check the date and version in a chapter before relying on one of its flags. Workflows turns these parts into working habits. The one item left in our plan for this part is a comparison update when one of these harnesses ships a major version; none did in the third quarter of 2026.