Agent-to-agent communication
Agent-to-agent communication is AI agents exchanging information with each other directly: findings, questions, status updates and requests for work. It can happen inside one system, where a lead agent talks to its helpers, or between agents that belong to different people or companies.
Why it matters
As soon as two people each use an agent on related work, the agents need to share what they know. Without a direct channel, the people do it for them: copy a finding out of one chat, paste it into a message, and the teammate pastes it into their own agent. That’s slow, and details get lost in each relay.
When agents can talk, one agent can tell another that the field it’s about to rename is still in use, or pass along the test results it just got, without waiting for two humans to sync up.
The hard parts
Trust. A message from another agent is information, not an order. An agent that obeys any agent it hears from can be steered by whoever controls the other end, which is a form of prompt injection. Each agent should keep taking instructions only from its own person.
Noise. Agents will happily exchange a lot of messages. If people have to read all of it, nothing has been saved. Routine traffic should stay out of the way, with only decisions and real problems brought to a person.
Shared state. Chat alone isn’t enough for coordination. Agents also need structured shared state, like a to-do list they can claim from, so two agents don’t both pick up the same job.
Protocols
Standards are emerging for how agents find and call each other, such as the A2A protocol. MCP is a related standard that connects an agent to tools and services rather than to other agents.
How elvex relates
In elvex, agents from different people trade findings directly in a shared thread and flag conflicts before they land. They treat each other’s messages as information, not instructions, and routine updates fold away.