
A2A agent-to-agent A rule for every handoff
Argent Strategic Intelligence governs agent-to-agent calls at the gateway: it decides which jobs each agent may ask another agent to do, refuses the rest by default, and signs a record of every call.
When agents hand work to other agents
Agent-to-agent (A2A) is the open protocol one AI agent uses to ask another agent to do a job: schedule a payment, look up an order, draft a reply.
The A2A specification leaves the decision to whoever runs the agent: "Authorization logic is implementation-specific" (A2A specification, section 7.5).
Argent Strategic Intelligence makes that decision at the gateway. It decides which jobs each agent may ask another agent to do, refuses the rest by default, and signs a record of every call.
Which jobs an agent may hand off is a rule you write, not a default you inherit.
What Argent checks on every agent-to-agent call
| Check | What happens |
|---|---|
| What is asked | Argent reads the A2A request (sending a message, streaming, checking or cancelling a task, and reading an agent's extended card, in the A2A 0.2 JSON-RPC format). A message it cannot read, or one that names no job, is refused. |
| Who is asking | The caller's identity comes from Argent's own sign-in layer, never from the message body. No identity, no call. |
| Which job | A job with no rule switched on for your tenant is refused. |
| Permission | Each rule names the one permission a caller must hold for that job, in the form scope:payments.schedule. The caller holds it exactly, or the call is refused. |
| If Argent cannot check | A policy engine fault or an unreadable rule refuses the call. |
| The answer | An allowed call goes to the peer agent, and its real reply comes back, marked as untrusted text. If the peer cannot be reached, the caller gets an error, never a made-up success. |
Only A2A traffic is examined. Everything else, MCP included, passes through to its own gate.
Agent cards and agent registries
In A2A, an agent publishes a card that says what it can do. The specification's standard place for it is `/.well-known/agent-card.json` on the agent's own host (A2A specification, section 8.2).
Argent reads those cards through its agent registries, under Integrations, Agent registries in the console:
- Embedded registry. The built-in default every tenant starts with. It needs no network and reflects your own A2A rules.
- Well-known registry. Give it an agent host's base address and Argent fetches that host's agent card from the standard path, falling back to the older
agent.jsonname only when the new one is not found. Argent never follows a path you type, so the address cannot be bent to fetch something else. - Live read on test. Switch on the live read option and Test fetches the card and shows the capabilities and skills the agent says it offers.
Cards tell your admins what an agent says it can do. What a caller may actually ask for is still decided by your rules.
How it connects
Agents send their A2A calls through Argent's gateway instead of straight to each other.
- An add-on you switch on. A2A Governance is an add-on, off until you turn it on for a tenant on the Add-ons page.
- Per tenant. Rules and agents are kept per tenant; one tenant's rules never reach another's agents.
- Watch first. The A2A gate has its own watch and enforce setting per tenant. In watch, a call that would be refused still goes through and the would-be refusal is recorded.
- Same identity as everything else. The caller is the same agent identity Argent already knows, tied to a human owner. See identity tethering.
What you configure
A SuperAdmin or TenantAdmin sets up A2A governance for a tenant.
- Switch on A2A Governance for the tenant on the Add-ons page.
- Write a rule per job on the A2A Governance page, Policy tab: the job name, the permission a caller needs (for example
scope:payments.schedule), and whether the rule is on. A job with no rule switched on stays refused. - Give the permission to the agents that should be able to ask for that job.
- Add an agent registry under Integrations, Agent registries, if you want the partner agent's card in the console.
- Watch, then enforce. Read the would-be refusals in watch mode, adjust the rules, then switch the A2A gate to Enforce. See watch before you enforce.
Every agent-to-agent call is on the signed record
Every allowed and refused A2A call is signed in Argent's audit ledger.
- The A2A Governance page's Lineage tab lists each call: who asked, which job, ALLOW or DENY, the policy that decided, and a fingerprint of the message instead of its text. It shows live refusals and a graph of which agents called which.
- Changes to the rules are signed before they take effect. If the record cannot be written, the change does not go live.
- Argent works out which permissions each agent actually used, day by day, so unused ones can be trimmed.
Hand the record to an auditor with one-click audit evidence. See audit and accountability.
Where this fits
- Capabilities: agent reach control (tools, agents and documents), prompt injection defense (replies from other agents marked as untrusted), identity tethering, owner offboarding, watch before you enforce.
- Goals: agent identity and ownership, threat defense, audit and accountability.
- Use case: controlling agent tool access.
- Agent frameworks that build multi-agent systems: LangGraph, CrewAI, OpenAI Agents SDK.
- Also in MCP and agent-to-agent: MCP servers, the internal MCP registry, Official MCP Registry, Docker MCP Catalog.
- Industries where agents hand off regulated work: financial services, public sector.
Browse all integrations.
Want this running on your own use cases?
Book a demo