HatchWorks

Everyone talks about the model. Almost nobody names the thing that makes it an agent. Why?

The harness is the runtime that wraps the model: the loop, the tool access, the context management, the memory, and the controls. It is the single most important piece of agent infrastructure that most teams never name out loud.

You cannot reason about your agents until you can name what runs them.

👇 Link in the comments

If AI now moves the exit multiple, why is so little of it showing up in measurable value?

Part of the answer is timing.

Defensible AI takes time to build and document, and the hold period is finite.

A portfolio company that starts now carries real, evidenced AI assets into its next process.

The ones that wait get diligenced against the ones that did not.

The window is open. It is also narrowing as more companies build the capability.

👇 Link in the comments

Using Claude for a job it was never built for?

Anthropic now ships three surfaces on the same model, and the names make them sound interchangeable. They are not.

Claude is the chat app for thinking with a human in the loop.
Claude Code is the developer agent for the terminal and git.
Cowork is the desktop agent for documents, data, and business workflows.

Match the surface to the work, and everything gets faster.

Which one are you reaching for most this week?

👇 Link in the comments