The First and Last Harness Is Us
Kinds of harnesses
The belt: different harnesses for different jobs
Not every harness is the same, for the same reason Batman doesn't carry the same gadgets to every fight. Here are the main kinds you'll run into.
The chat harness. The simplest belt: a set of instructions, a conversation window, and some guardrails. Most people's first experience of AI lives here. It's great for questions and drafts, and it does almost nothing on its own.
The tool harness. This wires the model into outside apps and data. Connector standards like the Model Context Protocol (MCP) are a big part of this: a common plug so any model can use any tool that supports it. This is the utility pouch.
The memory harness. This gives the model your world: your documents, your notes, your past conversations. It searches them and pulls in the right pieces when they matter. This is the Batcomputer.
The coding harness. Today's most mature agents live here. The model reads a codebase, edits files, runs the tests, sees what broke, and tries again. Code is the ideal training ground because the work can check itself: either the tests pass or they don't.
The computer-use harness. Lots of software has no clean way for a program to talk to it. So this harness lets the model look at a screen and click and type the way a person would. It's slower and clumsier, but it reaches everything.
The orchestration harness. One lead agent breaks a big job into pieces and hands them out to helper agents, then pulls the results together. This is the Bat-family: one strategist, many specialists.
The personal agent harness. Always on. It remembers you, runs on a schedule, and works across your accounts while you sleep. This is Alfred.
The safety harness. Sandboxes the agent can't escape, permissions it can't exceed, approval steps before anything irreversible, and logs of everything it did. Every other harness sits inside this one, whether the builder admits it or not.