Tagged tool use · show all articles
Four calling conventions, each dated to a real announcement: a manifest borrowed from web APIs, a schema built into the model call itself, incompatible framework wrappers, and a standard now governed by the vendors who once competed on incompatibility.
A model that only emits text cannot query a database or send an email. Four documented patterns cross that boundary, and each one answers a different question — none is a smaller version of another.
Grading a model once meant scoring a string against a reference answer. Grading an agent required borrowing a games lab's tools, discarding them, and rebuilding evaluation from scratch, in a documented, dated sequence of its own.
The failure is rarely the model. It is the loop with no edge, the handoff that drops what mattered, the fan-out that multiplies exposure, and the checkpoint nobody wired in before an irreversible action.
A working agent is a loop with a contract — observe, plan, act, reflect — plus explicit rules for what persists, what a tool call returns, and where a human has to say yes.
An agent can do exactly what its tool interfaces let it propose, no more. The schema, the granularity, the retry semantics and the error surface are not plumbing around the intelligence — they are the boundary of it.
Measuring an agent is not measuring a model with extra steps. The object under test is a route through an environment, and a single success indicator reads only its last term.