Two vendors named the same thing. The agent isn't the model. It's the harness

Microsoft and LangChain shipped the same idea in the same week, and they used the same word for it.

The word is harness.

Microsoft's Agent Framework at BUILD 2026 made the harness a first-class layer: context compaction, file memory, approval rules, and tracing. https://devblogs.microsoft.com/agent-framework/microsoft-agent-framework-at-build-2026-announce/

LangChain put it plainly. An agent is a model plus a harness. https://www.langchain.com/blog/how-to-build-a-custom-agent-harness

Most teams still treat the model as the product.

They swap one model for another, tune the prompt, and wait for the next release to fix what is broken.

But the gap between a demo and a production agent was never the model.

It is everything wrapped around the model.

The harness is the part that decides what the model sees, what it is allowed to do, and what happens when it fails:
• compacting context so the window does not rot
• persisting memory across steps
• retrying and recovering from tool errors
• gating risky actions behind approval

Microsoft measured one piece of this. CodeAct, which collapses a chain of tool calls into a single sandboxed program, cut latency 52.4% and tokens 63.9%.

That number did not come from a better model. It came from better scaffolding.

When two competing vendors converge on the same abstraction in the same week, the abstraction is the signal.

A better model makes a better demo.

A better harness makes a better system.