Agent2Agent is a protocol for one agent to discover and invoke another. It is the interop layer. MCP is how an agent calls tools and data. You usually want both.

Do we need A2A if we only have one agent framework?

Not on day one. It becomes useful the moment a second runtime appears — Agentforce plus CrewAI is the usual trigger.

How do you stop agents from calling anything they can see?

Capability cards, identity (often IndyKite or your IdP), and an allow-list. Discovery is not the same as permission.

Does delegation escalate privilege?

Not if on-behalf-of identity travels with the request. When agent A asks agent B, the user’s authority goes too, and hop budgets stop cyclic delegation. That is the runtime plane working as designed.

Salesforce MVP Hall of Fame · Certified Partner since 2010

Agents that can talk to other agents.

A2A is the Agent2Agent protocol: a standard way for one agent to discover and call another. We use it so a CrewAI invoice crew, a Google ADK helpdesk agent, and Salesforce Agentforce are not three silos. Mindcat implements the interop, the identity boundary, and the audit trail. The first week is in the org, reading what already exists.

Talk to an AI consultant

01Models · Claude, Gemini, OpenAI, Bedrock
02Agents · Agentforce, CrewAI, ADK, A2A
03Memory · Neo4j, Cognee, RAG
04Identity · IndyKite, SSO, entitlements
05Governance · policy, audit, human override

01 · MODELS

02 · AGENTS

03 · MEMORY

04 · IDENTITY

05 · GOVERNANCE

Hover to separate the layers

What we put in

The work on this layer.

The honest no: if the work is a software factory with a named delivery date, we are the wrong partner.

01A2A capability cards and discovery for each agent you want to exposeA2A capability cards and discovery for each agent you want to expose

02Auth and scope so an agent can only call what it was approved to callAuth and scope so an agent can only call what it was approved to call

03Logging that shows which agent asked, which agent answered, and what data movedLogging that shows which agent asked, which agent answered, and what data moved

04A written map of the agent graph your security team can reviewA written map of the agent graph your security team can review

Use cases

Where this shows up in the work.

These are the use cases we usually pair with A2A Protocol.

AI IT Support Transformation

Ticket routing and a knowledge desk. An agent on the repeats. A person on the exceptions.

CIO

AI Customer Service Transformation

Cases, knowledge, and the channel the customer already uses. An agent on the repeats. A person on the exceptions.

COO

Supply chain

Inventory and orders between ERP and Salesforce when that is the job. Real-time or batch from the desk.

CTO

Questions

What we actually say.

What is A2A?: Agent2Agent is a protocol for one agent to discover and invoke another. It is the interop layer. MCP is how an agent calls tools and data. You usually want both.
Do we need A2A if we only have one agent framework?: Not on day one. It becomes useful the moment a second runtime appears — Agentforce plus CrewAI is the usual trigger.
How do you stop agents from calling anything they can see?: Capability cards, identity (often IndyKite or your IdP), and an allow-list. Discovery is not the same as permission.
Does delegation escalate privilege?: Not if on-behalf-of identity travels with the request. When agent A asks agent B, the user’s authority goes too, and hop budgets stop cyclic delegation. That is the runtime plane working as designed.

The brief

More than one agent runtime already?

We will map the agents you have, what they should be allowed to call, and whether A2A is the right glue.