Vol. V · The Harness · V.1
What a Harness Is
A harness is the opinionated product around the loop: workspace, tools, skills, subagents, permissions, UI. Frameworks are libraries. Harnesses are how an engineer actually works with a model.
7 min read
Doctrine
- If you are building a coding or research agent, start from a harness shape, not from a blank ReAct loop.
- The common core: a filesystem, a planner, tool permissions, a skill loader, a subagent spawner, a sandbox.
- Replace pieces. Do not fork the universe. Deep Agents' pitch is exactly this.
The convergence
Claude Code, Codex, OpenCode, Cline, Gemini CLI, OpenHands, Aider, Deep Agents, DeerFlow — different brands, same skeleton. A workspace the model can read/write. A way to run commands under policy. Memory that survives sessions. Skills that load on demand. Subagents for isolation. A human approval gate on dangerous tools.
ByteDance DeerFlow adds a message gateway (meet users where they are — Factor 11). LangChain Deep Agents adds LangGraph durability. OpenHands adds a full computer. Aider adds git as the transaction log.
Anti-patterns
- Rebuilding Claude Code poorly in a weekend without permissions or a workspace.
- Calling a chat UI with function calling a harness.