Integrations Concept

Integrations are approved external capabilities that agents can use during a run or thread.

Integrations give agents access to systems outside the model: search, work management tools, communication tools, databases, and other services.

Capability Boundaries

Each integration should be attached intentionally. Before enabling one, decide:

  • What the agent can read.
  • What the agent can write.
  • Which users are allowed to trigger it.
  • Whether it is safe in public chatbot contexts.

Public Agent Defaults

For public agents, start with read-only integrations. Add write-capable actions only when the user experience, permissions, and audit trail are clear.

User-Facing Assistants

Your users do not need to know which integration an agent used unless your product wants to show that. What matters is that the assistant has the right permissions, gives clear answers, and handles blocked actions gracefully.