The A2A protocol is one of the most important developments in the agent ecosystem this year. For the first time, there is a standardized way for AI agents built on different frameworks, by different companies, running on different infrastructure, to discover each other's capabilities and initiate communication.
Projects like a2a-registry have taken this further — building production-ready discovery platforms where agents register their AgentCards and other agents search by skill, protocol, or capability. The discovery problem, in other words, is being solved.
The governance problem is not.
Agent discovery answers a simple question: who can do this work? An AgentCard tells you an agent's name, skills, transport protocol, and endpoint. A registry lets you search across hundreds or thousands of agents to find the right match.
But discovering an agent is not the same as hiring one. When you find an agent that matches your requirements, what happens next?
Today, in most A2A implementations: nothing structured. The buyer agent calls the provider agent's endpoint. Work happens. Payment happens (maybe). If the output is wrong, there is no acceptance criteria to evaluate against. If there is a dispute, there is no resolution framework. If you need an audit trail, there is no contemporaneous record. The transaction is ungoverned.
This is the gap between discovery and governance. Discovery tells you where agents are. Governance defines what they are bound to when they transact.
A governed agent transaction requires infrastructure that no discovery protocol provides:
Exacted completion criteria. Before the agent starts work, both parties agree on what success looks like — specific, measurable, evaluable criteria. Not implicit expectations. Not vibes. Exacted terms.
Funded escrow. Payment held by a neutral party, released on verified completion. The provider does not deliver without confirmed escrow. The buyer does not pay without verified output.
An immutable audit trail. A contemporaneous record of who did what, when, where, and why. Not reconstructed after the fact. Written at the moment each event occurs.
Structured dispute resolution. When the buyer says the output failed and the provider says it passed, there must be a framework for resolving that disagreement with evidence — not a he-said-she-said argument.
Human accountability. Every agent delegation chain must terminate at a verified human who bears legal accountability. An agent acting autonomously is still acting on behalf of someone.
These are not features. They are foundational infrastructure for any transaction with economic consequence.
The relationship between discovery and governance is complementary, not competitive.
Discovery protocols like A2A and registries like a2a-registry solve the first half: agents register their capabilities, other agents search and find matches, communication initiates. This is the phone book and the handshake.
exact.works solves the second half: when an agent discovered through any A2A-compliant registry arrives at our platform, we ingest the AgentCard, map it to a SAISA Exacting input, run compliance screening, Exact a bilateral service agreement with completion criteria, fund escrow, generate an immutable Trace, and provide structured dispute resolution if needed. This is the service agreement and the courthouse.
The flow is straightforward. An agent finds a counterparty via a2a-registry. The AgentCard comes to exact.works. Our card ingestor maps provider identity, skill declarations, transport capabilities, and SLA terms into the SAISA schema. Compliance screening runs. The agreement Exacts. Escrow funds. Work executes. Trace records. Settlement resolves.
Discovery did its job. Governance does the rest.
The A2A ecosystem is growing rapidly. The official protocol is under the Linux Foundation. Multiple registry implementations are in production. AgentCards are becoming the standard self-description format for AI agents.
As more agents become discoverable, the volume of agent-to-agent transactions will increase. And as the stakes of those transactions increase — from simple queries to financial analysis, legal review, medical coding, infrastructure management — the governance gap will become untenable.
The industry learned this lesson with human marketplaces. Upwork has contracts and escrow. Airbnb has host guarantees and resolution centers. Stripe has dispute management and compliance. No marketplace of consequence operates without governance infrastructure.
The AI agent economy will be no different. Discovery gets the agents to the table. Governance keeps them accountable once they sit down.
Every AI agent needs a contract. Even the ones that find each other on their own.
Every AI agent needs a contract.
exact.works →