Series
Agent harness design.
The product concerns that decide whether an agent is reliable enough to ship.
An agent harness is the layer between the model and the world: the loop, the tools, the memory, the permissions, the trace. The model sets the ceiling on what an agent can do. The harness sets the floor on how reliably it does it, and the floor is what decides whether you can put it in front of real users.
Almost every decision in the harness that looks like an engineering choice is really a product choice about behaviour, risk, and trust. This series takes the floor apart, one concern at a time. Each piece stands on its own; together they are a short book.
- I
The opening case. Production removes the supports your laptop quietly provided, one at a time, and the model was never the hard part.
- II The authority surface Published
What the agent may do alone, what needs a human, and what it must never do. Drawing that line, action by action, is a design, not a setting.
Related: Ironies of AI automation
- III The action interface Planned
The tools you expose are the agent's capability boundary. Designing them for a caller that is not a person: granularity, error semantics, and what is safe to repeat.
Related: When the caller cannot ask
Each essay also publishes as an A4 PDF. Further parts are in progress, and the complete set will compile into a short book.