A Harness Ships Ready to Run. A Framework Still Has to Be Wired
Claude Code and Codex are pre-wired loops. LangGraph remains a kit for branching, durable workflows.
Direct answer
Claude Code and Codex are productized harnesses for one vendor's models, while LangGraph is a framework you assemble when the job needs branching and checkpoints.
By mid-2026 the word harness had a stable meaning. A harness is a pre-wired agent loop: tools, permission checks, and session state, already assembled, pointed at a task. A framework is the opposite shape. It hands you the parts and expects you to wire the loop yourself.
That line puts Claude Code and Codex on one side and LangChain or LangGraph on the other. Claude Code is a TypeScript harness bound to Anthropic models. Codex is the OpenAI counterpart, with a Rust and TypeScript CLI under Apache-2.0 and a cloud sandbox that can run the programs the agent writes. A 2026 comparison treats the choice as model and stack lock-in more than a gap in raw capability.
One loop is not an organization chart
The same comparison keeps a second distinction. The harness is how one agent runs. Orchestration is how several agents share a job: a parent that delegates, a fan-out across search and tests, or a handoff to a specialist. LangGraph belongs there. Its nodes and conditional edges fit workflows that must branch and resume after a crash. It is not a substitute for the productized loop inside Claude Code or Codex.
Pydantic's June 2026 release is the tell that the layer is real. Version 2.0.0 is described as harness-first: a capability bundles tools, hooks, instructions, and model settings. By 19 August 2026 its documentation listed more than fifty of those capabilities. A framework absorbing harness ideas does not erase the product boundary. It shows teams now expect the loop to arrive already designed.
Use a product harness when the work is sustained coding on one model family. Reach for a graph framework when you need an audit trail across branches. Mixing the two names is how architecture advice goes wrong.
FAQ
- What is the difference between a harness and a framework?
- A harness is a pre-wired agent loop shipped as a product. A framework gives you parts and leaves the loop, tools, and permissions for you to assemble.
- When does LangGraph still matter?
- When one loop is not enough: auditable branches, crash recovery, or several agents that have to be coordinated.