Connect A2A Agents
Your employees can delegate tasks to external AI agents using Google's Agent-to-Agent (A2A) protocol. If another agent exposes an A2A endpoint, your employees can discover its capabilities and send it work, just like calling any other tool.
TL;DR
Add an external AI agent's URL in any employee's Tools tab and the platform discovers what the agent can do. Your employees can then delegate tasks to it. This is for advanced users who want their employees to collaborate with AI agents running outside the platform.
What Is A2A?
A2A (Agent-to-Agent) is Google's open protocol for AI agents to communicate with each other. While MCP connects agents to tools (structured operations like "search" or "create"), A2A connects agents to other agents that can reason, plan multi-step work, and make decisions.
An A2A agent publishes an Agent Card describing who it is and what it can do. Other agents discover it and send tasks.
How It Works
When you add an A2A agent, the platform:
- Fetches the Agent Card from the agent's URL
- Reads the skills list (names, descriptions) from the card
- Creates a tool record with each skill as a discovered tool
- Assigns it to all employees (current employee enabled, others disabled)
- At runtime, each skill becomes a callable tool that sends a task to the external agent
- The external agent processes the task and returns a response
Each skill from the Agent Card becomes an individual tool your employee can call. If the agent has 3 skills, the employee sees 3 tools. The employee doesn't need to know it's talking to another agent. It looks and works like any other tool.
When to Use This
| Scenario | Example |
|---|---|
| Specialized analysis | Your marketing employee delegates data analysis to a specialized analytics agent |
| Expert escalation | Your support employee escalates complex technical issues to a domain-expert agent |
| Research delegation | Your sales employee asks a research agent to compile competitive intelligence |
| Content review | Your content employee sends drafts to an editorial agent for style and grammar review |
| Candidate screening | Your HR employee delegates candidate screening to a specialized recruiting agent |
A2A vs MCP
| A2A | MCP | |
|---|---|---|
| Connects to | Other AI agents | Tools |
| Best for | Tasks requiring judgment, reasoning, multi-step planning | Structured operations with clear inputs and outputs |
| Input | Free-text task description | Structured parameters |
| The other side | Thinks and reasons about the task | Executes a specific operation |
| Example | "Analyze this data and give me insights" | "Run this SQL query" |
This page covers connecting your employees TO external A2A agents. To expose your employees AS A2A agents for external systems, see Integrate via A2A in the Channels and APIs section.
Adding an A2A Agent
- Open any employee's Tools tab
- Scroll to the Integrations section
- Find the A2A Agents group card and click the "Add A2A Agent" pill button
- A drawer slides open with the connection form
- Fill in:
- Name: a label you'll recognize (e.g., "Research Agent", "Data Analyst")
- Agent URL: the A2A agent's base URL (e.g.,
https://agent.example.com/api/a2a/) - Auth Header: optional, if the agent requires authentication (e.g.,
Bearer sk-...)
- Click Connect
- The platform fetches the Agent Card and shows discovered skills
What You See After Connecting
Once added, the A2A agent appears as a card inside the A2A Agents group in the Integrations section. Each card shows:
- Discovered skills: expandable list showing every skill from the Agent Card, with descriptions
- Tool Rules: a text field where you can add custom instructions that get appended to every skill's description at runtime
- Enable/disable toggle: per-employee control
Auto-Assignment
A2A agents follow the same pattern as MCP servers. When you add an agent:
- It gets assigned to all employees automatically
- The employee you're configuring has it enabled by default
- All other employees have it disabled by default
Managing A2A Agents
Enable/Disable
Toggle the switch on the A2A agent card to enable or disable it for a specific employee.
Refresh Skills
If the external agent adds new skills or updates its Agent Card, click the Refresh icon on the card. The platform re-fetches the card and updates the discovered skills list.
Remove
Delete the card to archive the A2A agent. It will be removed from all employees.
Tool Rules
Add custom instructions in the Tool Rules field. These get appended to every discovered skill's description at runtime.
Requirements
- The external agent must support Google's A2A protocol
- The agent must publish an Agent Card (usually at
/.well-known/agent.json) - The agent must be reachable from the internet
Good to Know
- Stateless connections. The platform connects to the A2A agent per tool call. No persistent sessions
- Cached skills. Skill descriptions are cached from the Agent Card. The agent only needs to be reachable when executing tasks
- Same billing. A2A tool calls are part of the employee's normal workflow. Credits are consumed for the LLM reasoning, not for the task delegation itself
- Error handling. If the external agent is down, the employee sees an error and can retry or work around it
- Two-way A2A. Your employees can both call external A2A agents (this feature) and be called by external agents (see the A2A Protocol guide in Channels and APIs)
Troubleshooting
- Agent Card not found. The platform fetches
/.well-known/agent.jsonfrom the base URL. Make sure the agent publishes its card at that path, or that the URL you entered points directly to the card - No skills discovered. The Agent Card exists but has an empty skills list. The external agent needs to list its capabilities in the card
- Auth errors. Double-check the Auth Header value. It should be the full header value (e.g.,
Bearer sk-...), not just the token - Task execution failing. The agent was reachable during discovery but may be down now. Check that the agent's task endpoint is operational
Frequently Asked Questions
Q: What's the difference between A2A and MCP? A: MCP connects your employee to tools that execute specific operations (search, create, query). A2A connects your employee to other AI agents that can think, plan, and handle complex multi-step tasks. Use MCP when the task is structured, A2A when the task requires judgment.
Q: Does the external agent see my employee's full conversation? A: No. Your employee sends only the specific task description to the external agent. The agent receives the task, processes it, and returns a result. It does not have access to your employee's memory, history, or other tools.
Q: Can my employee and the external agent go back and forth? A: Currently, it's a single request-response. Your employee sends a task, the external agent returns a result. Multi-turn negotiation between agents is not supported yet.
Q: Do I need to build my own A2A agent to use this? A: No. You can connect any third-party A2A agent that's publicly available. Building your own requires development skills, but connecting an existing one is as simple as pasting a URL.
Q: How are A2A calls billed? A: You pay standard credits for your employee's thinking time. The A2A task delegation itself has no additional cost from our side. The external agent provider may have their own pricing.