Seeing as an Agent

Every tool has two users: the human who triggers the agent, and the model that decides how to call it. Design for both.

The model has no memory across turns, no outside world except through tools, non-deterministic output, and a training cutoff. Designing from that existence means asking:

• What tools fit its abilities, not just the problem's complexity?
• Are descriptions internally consistent? Is the output format natural given how it was trained?
• What reduces its uncertainty about when and how to use a tool?
• What helps at one capability level may constrain the next — what scaffolding becomes a cage?
• Where are its edges, and how can the harness serve rather than just constrain it?

Debugging: trace from the model's side first

When a tool call goes wrong, resist the urge to fix the code immediately. Instead: