A2A and Multi-Agent Protocols
MCP standardizes how agents connect to tools and data. Agent-to-agent protocols standardize how agents discover, delegate, and coordinate with other agents.
Think of MCP as "agent to tool" and A2A-style protocols as "agent to agent." Production systems often need both.
Why multi-agent systems exist
A single agent can become too broad. Multi-agent systems split work into specialized roles:
- researcher
- planner
- coder
- reviewer
- data analyst
- compliance checker
- customer support triage
Common orchestration patterns
| Pattern | How it works | Best for |
|---|---|---|
| Supervisor | one planner delegates to specialists | controlled workflows |
| Sequential | each agent hands off to the next | document pipelines |
| Parallel | agents work independently then merge | research and review |
| Debate | agents critique each other | high-stakes analysis |
| Marketplace | agents advertise capabilities | open ecosystems |
What protocols add
Protocols can define:
- agent identity
- capabilities or skills
- input/output schemas
- streaming status
- task IDs
- authentication
- cancellation
- human approval points
A2A vs MCP
| Need | Protocol style |
|---|---|
| Read a database | MCP server |
| Search a filesystem | MCP server |
| Ask a specialist agent to review code | A2A-style agent handoff |
| Delegate research to another runtime | A2A-style task |
| Call a calculator | MCP tool |
Safety rules
- Do not let agents discover unlimited peers automatically in high-risk systems.
- Keep agent permissions narrower than user permissions.
- Log every handoff and tool call.
- Set budgets for tokens, wall-clock time, and tool calls.
- Require approval for actions that affect external systems.
Knowledge check
Q1: What is the difference between MCP and A2A-style protocols?
MCP connects agents to tools/data; A2A-style protocols connect agents to other agents.
Q2: What is the biggest multi-agent risk?
Uncontrolled delegation that amplifies cost, latency, or unsafe actions.