Integrates external APIs for enriched task execution, information retrieval, and actions.
Gives an agent a registry of callable APIs and lets it select and invoke them. The mechanism behind most useful agentic behaviour, and the surface where the serious security questions live.
Tasks requiring information or actions across several systems, where the combination needed varies by case.
A fixed sequence of two or three calls. Write the code.
Read and write capabilities must be separate tool sets with separate authorisation. An agent that can read freely and write only through an approval gate is a fundamentally different risk profile from one with a single permission scope.
Over-broad credentials. The agent is given a token with everything the integration could ever need, so a prompt injection or a reasoning error has the full surface available.
An investigation agent that queries balances, match records and prior resolutions freely, and can draft but not post a correcting entry.