Agents & Tools

A model's tool call is text until the harness makes something of it: parse it, look the tool up, check every argument against the schema, run it, and hand the result back. Every step can say no.

An agent is a loop: the model proposes, the harness runs the tool and appends what came back, and round it goes until an answer, a turn cap, a token budget, or too many errors. The model is a script.

A call is a line of JSON with a method, params, and an id; the reply comes back under the same id, in any order. Leave the id out and nothing comes back at all. The wire is someone else's problem.

JSON-RPC with a lifecycle and a vocabulary: initialize negotiates version and capabilities, then the client lists and calls tools, reads resources, and gets prompts, asking only for what was offered.