Agents
Updated on 11 November 2026.
An agent is a target that callers reach through the gateway. Go to Settings > AI Gateway > A2A Gateway. The Agents tab shows totals for the period and one card per agent.
Each agent card shows its gateway URL, A2A version, bindings, outbound auth, pinned card version and traffic. Use Inspect to open its logs or card versions, and Configure to jump to a section of the agent page. The switch turns the agent on or off for every caller.
Add an agent
Click Add agent. The wizard has four steps: Source, Connection & auth, Review card and Access & defaults. The Quick start walks through it.
Outbound auth
How the gateway authenticates to the agent sets the credential the gateway adds to every call. The caller's key is never forwarded. Secrets are write-only: once saved, they are never shown again.
There is also an option that forwards the caller's verified end-user token as is. It is not recommended; use token exchange instead.
Access & defaults
The agent page
Click an agent's name to open its page. The links under the title jump to each section.
- Endpoints: the agent's gateway URL (for both JSON-RPC and HTTP+JSON), its agent card URL, and a
curlexample you can copy. On a skill-routed agent, the example includes the skill. - Card versions: the pinned card and any versions waiting for review.
- Skills: the pinned card's skills (disabled ones are marked), the agent's own interfaces and the signature check.
- Callers with access: every caller with a grant on this agent. Manage callers opens the callers list.
- Test: run a call as a chosen caller. See Test an agent.
- Settings: everything from the wizard, plus timeout, hop depth, limits and re-entry. See Agent settings.
- Audit log: who changed what on this agent.
The ... menu has Refresh card (check the card now) and Delete agent.
Card changes
Callers always get the card version you approved, called the pinned card. The gateway never fetches a card while handling a call.
The gateway checks each card-URL agent's card every 15 minutes. When the card changes, the new version waits under Card versions with a summary of what changed, for example "1 skill added". Click Approve to pin it or Reject to keep the current one.
Choose what happens while a change waits with When the agent publishes a new card:
A new skill is never callable until you approve the version that adds it. Callers with All current and future skills then get it; callers granted a list of skills do not.
If the gateway cannot read an agent's card, the agent stays usable on its pinned card, and the Agents tab lists it under Needs attention.
Test an agent
Use Test this agent to check a caller's access before you share an agent.
- Pick a mode:
- Simulate checks access, limits and data protection. The agent is not contacted and nothing is logged.
- Send a real message calls the agent for real, so it can have side effects such as payments or tickets. It needs a separate permission and is logged as an admin test.
- Choose the caller under As caller, the Skill, and type a Message.
- Click Run simulation (or send the message).
The result shows each check: caller, grant, method, skill, the agent endpoint the gateway would use, limits and data protection. View raw shows the full response.
Agent settings
Skills you untick on the Review card step are disabled: hidden from every caller and refused.
Wrapped agents
A wrapped agent is one that does not speak A2A. The gateway acts as the A2A server for it: callers see a normal A2A agent, and the gateway calls the platform for them.
- In Add agent, choose Wrapped (non-A2A) agent and pick the Platform.
- Fill in the Connection fields. The wizard lists what you need and the minimum permissions for the platform you picked.
- Under Agent card, enter the name, description and skills callers will see.
- Finish the wizard as for any agent.
The HTTP option has three presets:
Good to know:
- Wrapped agents are always skill-routed. A caller granted one skill can leave the skill out; the gateway fills it in.
- Each skill can use its own platform settings.
- Wrapped agents accept text and data parts. File parts are refused.
- Streaming works on AgentCore, Vertex AI, Azure AI Foundry and the two OpenAI presets. Cancel works on Azure AI Foundry, and on AgentCore as best effort.