Post

The more I work on harnesses, the more I’ve come to dislike the term and metaphor, for two reasons. Note, terminology here is roughly: model (self-explanatory) + harness (software around model, can simply be agentic loop) = agent. This may not be universal but it’s common enough. So the first

reason is the more practical one: the metaphor is brittle. While most people just interact with a coding harness, they’re also extensible - MCP servers, skills, scripts, etc. can be bolted on to any decent coding harness. These are increasingly packaged separately, as a harness for your harness.

IRL you don’t harness a harness, that’s nonsense. The metaphor is also broken definitionally. A harness is used to direct, yes, but also (and more core to its function) hold something back. Agent harnesses don’t hold anything back! They give models their agency and augment their abilities, unlike

real harnesses. I keep coming back to the mech metaphor for these tools, because it illustrates well how they augment (and are directed by) humans, but even with that I don’t know if we have a term for reusable augmentation pieces that can augment the “core” (model) or the mech itself (coding agent)

I'd argue that a harness does place some constraints on the model though as in which tool calls are available to the model? It does hold things back as well (context compaction and management)

The base state of the model is that nothing is available to it unless provided to it (by the harness), and that there's no spontaneous/internal activity to be constrained from just the model.

One other thing comes to mind is that the common usage of harness in other contexts in software (e.g. test harness) mostly describes the connection and scaffolding components around the core machine, rather than a direct metaphor for like a riding harness on an animal