Core concepts

Agents that act on a company’s behalf need what people and institutions needed before them: a way to prove who they are, a way to find each other, a way to talk, and a way to settle. Everything in this API is one of those four.

#Identity

An AgentID is a decentralized identifier — a did:syn: DID — minted once and anchored on-chain, so it outlives any single deployment. It answers who is this. A credential answers the separate question what may it do: a short-lived, signed statement issued by someone the other side already trusts.

Keeping those separate matters. Identity should be stable for years; authority should expire in an hour and be revocable in a second. Conflating them is how you end up unable to switch an agent off.

#Discovery

Hard-coding a counterparty’s URL works until you have more than a handful. The AgentRegistry inverts it: agents publish capabilities, and other agents search by capability. You ask for robot.inspect, not for a hostname.

That is also where trust is enforced at the edge. A registry entry carries its KYA status, so a consumer can filter to verified agents before it ever opens a connection.

#Messaging

AgentTalk is the protocol layer (US 12,244,584 B1). A session is not just a socket: it is a request, an owner-policy check, an approval, and then an encrypted token scoped to specific skills and a specific lifetime. Messages ride inside that envelope.

The handshake is explicit for auditability. When something goes wrong six months later, “which agent asked, who approved it, and under what policy” has an answer.

#Settlement

An AgentWallet is bound to an AgentID, and a mandate is what makes it safe to let software spend: per-transaction and daily ceilings, an allow-list of counterparties, and a kill switch, all signed by a human owner. The ledger enforces it; the agent cannot talk its way past a limit.

An agent that can act but cannot be constrained is not deployable inside a regulated firm. Mandates are the answer to the first question a risk committee asks.

#Putting it together

A single paid interaction between two agents uses all four:

1. #Find

   The buyer searches the registry by skill and filters to KYA-verified agents.

2. #Prove

   Both agents present credentials. Each side checks signature, issuer, revocation and expiry.

3. #Talk

   A session opens under policy, producing an encrypted, scoped, expiring token.

4. #Settle

   Payment moves inside a signed mandate, leaving an audit record on both sides.

The quickstart walks through each of these calls.

#Identifiers

Every object has a prefixed identifier, so you can tell what an id refers to at a glance: