Glossary
Agent-to-agent (A2A)
Agent-to-agent refers to AI agents communicating directly with one another — delegating tasks, asking questions and returning results — rather than everything routing through a person.
Two scopes
Inside one deployment, agent-to-agent is a platform feature: your agents message each other under permissions you set. Across organisations, it is a protocol question — your agent talking to someone else's, which needs a shared format for discovering what an agent can do and addressing it safely.
The security rule
An agent should only be able to reach parties you named in advance, never an arbitrary address handed to it at runtime. The same applies to credentials: an agent following a link to an unexpected endpoint must not carry your authentication there.
Visibility is not permission
Telling an agent that a peer exists is a separate decision from allowing it to message that peer. An agent aware of many colleagues will try to involve them.
Related terms
FAQ
What is A2A?
Agent-to-agent communication: AI agents delegating tasks, asking questions and returning results directly to each other rather than routing everything through a person. Inside one deployment it is a platform feature governed by permissions; across organisations it is a protocol question.
Is agent-to-agent communication safe?
It depends on addressing. An agent should only reach parties named in advance, never an arbitrary address given at runtime, and it must not carry your credentials to an unexpected endpoint. Visibility should also stay separate from permission — knowing a peer exists is not authorisation to message it.