Tool calling
A model invoking external functions — search, database queries, file writes — by emitting structured calls its runtime executes. Tool calling is what turns language models from advisors into actors.
The moment models became actors
Before tool calling, a model could only describe what should happen. With it, the model emits a structured request that its runtime executes against a real system, then feeds the result back so reasoning can continue. That loop — decide, call, observe, decide again — is the entire mechanical basis of agent behaviour.
What the runtime is responsible for
Validating arguments against the schema, enforcing whether this agent may call this tool at all, executing it, handling failure without pretending success, and returning a result the model can actually use. Models propose; runtimes dispose. The quality of that layer decides whether agents are dependable.
Designing tools models use well
Narrow tools with unambiguous names and small argument sets outperform flexible ones. Errors should be actionable — "no record matched that ID" teaches the model something, "500" does not. And anything irreversible deserves a confirmation step outside the model’s discretion.
In AGNT
Every AGNT tool call is executed by the runtime, logged, and shown in the run’s receipt.
Tool calling — common questions
Is tool calling the same as function calling?
Nearly. Function calling names the model-side mechanism that emits the structured invocation; tool calling usually describes the whole loop including execution and the result coming back.
What happens when a tool call fails?
In AGNT the failure is surfaced rather than swallowed: it appears in the receipt, and the agent can retry, choose a different approach or escalate.
Can I limit which tools an agent may call?
Yes. Toolkits are assigned per agent, so an agent simply cannot invoke what it was not given.