A single AI agent can automate a task. But real business processes span systems, teams, and organizations. Your sales automation needs data from a market research agent. Your compliance workflow needs a verdict from a legal analysis agent. Your support pipeline needs to invoke a billing system agent to issue refunds.
The Agent-to-Agent (A2A) protocol makes this possible. JieGou now supports A2A in both directions: invoke external agents from your workflows, and expose your own recipes and workflows as agents that others can call.
What is A2A?
A2A is a protocol for AI agents to discover, authenticate with, and invoke each other. It’s built on a simple concept: agents publish a description of their capabilities at a well-known URL, and other agents use that description to send them tasks.
An agent’s capabilities are described in an Agent Card — a JSON document published at /.well-known/agent.json. The card lists the agent’s skills, input/output schemas, authentication requirements, and endpoint URL. Think of it as a machine-readable API spec, but for AI agents instead of REST endpoints.
Invoking external agents from workflows
JieGou adds a new workflow step type: A2A Step. Drop it into a workflow, point it at a registered agent, map your workflow data to the agent’s input schema, and the step invokes the external agent during execution.
The execution flow:
-
Discovery. Register an external agent by providing its Agent Card URL. JieGou fetches the card, validates it, and stores the agent’s capabilities in a registry. A health check confirms the agent is reachable (10-second timeout).
-
Configuration. In the workflow editor, add an A2A step. Select the registered agent, choose a skill, and map input fields from previous step outputs to the agent’s expected inputs.
-
Execution. When the workflow runs, the A2A step sends a task request to the external agent’s endpoint. JieGou polls for the result with configurable timeout and retry settings.
-
Result handling. The agent’s output is stored in the workflow context, available to downstream steps just like any other step output. If the agent fails, the step follows your workflow’s error handling policy — retry, skip, or halt.
A2A steps integrate with everything else in the workflow engine: approval gates, checkpoints, parallel branches, conditional logic, and error handling all work as expected.
Exposing your recipes as agents
The other direction is just as important. Any recipe or workflow in JieGou can be published as an A2A-compatible agent.
When you create a hosted agent, JieGou:
- Generates an API key (returned once in plaintext, stored as a SHA-256 hash)
- Publishes an Agent Card that maps your recipe’s input/output schemas to A2A skill definitions
- Creates an endpoint that accepts A2A task requests
External systems can discover your agent via its Agent Card, authenticate with the API key, and submit tasks. JieGou executes the underlying recipe or workflow and returns the result in A2A format.
This means your carefully tuned, tested, version-controlled recipes become callable services — without building a separate API, writing deployment scripts, or managing infrastructure.
Trust and authentication
Not all agents are equal. JieGou assigns a trust level to each registered agent:
- Verified — Approved by JieGou, vetted for reliability
- Community — Registered by other JieGou users, peer-validated
- Self-hosted — Agents you run yourself
- Unknown — Unverified external agents
Authentication supports API keys, bearer tokens, and OAuth2. Credentials are encrypted at rest and decrypted only at invocation time.
Audit trail
Every A2A invocation — inbound and outbound — is logged with:
- Direction — Whether JieGou invoked an external agent or received a task
- Agent and task IDs — For cross-system correlation
- Input and output previews — Truncated for privacy, sufficient for debugging
- Source metadata — IP address, user agent, caller identity
- Timing — Duration in milliseconds
- Status — Submitted, working, completed, failed, or canceled
- Workflow link — If the invocation was part of a workflow, a direct link to the workflow run
Audit logs are queryable and filterable by account, direction, status, and time range. When an external agent fails, you can trace the exact request and response without guessing what went wrong.
When to use A2A
A2A is most valuable when:
-
You need capabilities you don’t own. Your company has an internal ML model for fraud detection. Expose it as an A2A agent, and your JieGou workflows can invoke it natively.
-
You’re building a multi-agent system. Different teams own different automations. A2A lets them compose without tight coupling — the sales team’s lead scoring agent can feed into the marketing team’s campaign agent.
-
You want to offer your automations as a service. Publish your best recipes as hosted agents. Partners or other departments can invoke them without needing JieGou accounts.
-
You’re integrating with third-party agent platforms. Any platform that supports the A2A protocol can interoperate with JieGou — no custom integration required.
Availability
A2A protocol support — both inbound and outbound — is available on Enterprise plans. Hosted agent creation, external agent registration, and the full audit trail are included. Learn more about enterprise features or talk to sales.