Communication Protocols
Agents that can't speak the same language aren't a team. They're strangers shouting into the void.
Learning Objectives
- Implement MCP tool discovery and invocation so agents can use tools exposed by external servers
- Build an A2A agent card and task endpoint that allows one agent to delegate work to another over HTTP
- Compare MCP (tool access), A2A (agent-to-agent), ACP (enterprise audit), and ANP (decentralized trust) and explain which protocol solves which problem
- Wire multiple protocols together in a single system where agents discover tools via MCP and delegate tasks via A2A
The Problem
You split your system into multiple agents: a researcher, a coder, a reviewer. Each is good at its own job, and now they have to talk to each other.
The obvious first attempt is to pass strings around. The researcher returns a blob of text and the coder parses it however it can. That works until the coder misreads a research summary, or two agents deadlock waiting on each other, or you need agents built by different teams to collaborate. At that point "just pass strings" falls apart.
Without a shared contract for how agents exchange information, a multi-agent system is fragile, hard to audit, and stuck at the handful of agents you personally wrote.
The ecosystem has answered with four protocols, each covering a different slice of the problem:
- MCP for tool access
- A2A for agent-to-agent collaboration
- ACP for enterprise auditability
- ANP for decentralized identity and trust
This lesson reads the real wire formats from each spec, builds working implementations, and connects the protocols into one system.
The Concept
The Protocol Landscape
Think of these four protocols as layers, each addressing a different question:
They aren't competitors. Each one answers a different question at a different level.
MCP (Recap)
MCP is covered in depth in Phase 13. Quick recap: MCP standardizes how an LLM connects to external tools and data sources. It's a client-server protocol where the agent (client) discovers and calls tools exposed by a server.
MCP is agent-to-tool communication. It doesn't help agents talk to each other.
A2A (Agent2Agent Protocol)
Created by: Google (now under Linux Foundation as lf.a2a.v1)
Spec version: 1.0.0
Problem: How do autonomous agents collaborate, negotiate, and delegate tasks to each other?
A2A is the protocol for peer-to-peer agent collaboration. Where MCP connects an agent to tools, A2A connects an agent to other agents. Each agent publishes an Agent Card at a well-known URL, and other agents discover, negotiate with, and delegate tasks to it.
How A2A Works
The Real Agent Card
Here is what an A2A Agent Card looks like. It is served at GET /.well-known/agent-card.json:
{
"name": "Research Agent",
"description": "Searches documentation and summarizes findings",
"version": "1.0.0",
"supportedInterfaces": [
{
"url": "https://research-agent.example.com/a2a/v1",
"protocolBinding": "JSONRPC",
"protocolVersion": "1.0"
},
{
"url": "https://research-agent.example.com/a2a/rest",
"protocolBinding": "HTTP+JSON",
"protocolVersion": "1.0"
}
],
"provider": {
"organization": "Your Company",
"url": "https://example.com"
},
"capabilities": {
"streaming": true,
"pushNotifications": false
},
"defaultInputModes": ["text/plain", "application/json"],
"defaultOutputModes": ["text/plain", "application/json"],
"skills": [
{
"id": "web-research",
"name": "Web Research",
"description": "Searches the web and synthesizes findings",
"tags": ["research", "search", "summarization"],
"examples": ["Research the latest changes in React 19"]
},
{
"id": "doc-analysis",
"name": "Documentation Analysis",
"description": "Reads and analyzes technical documentation",
"tags": ["docs", "analysis"],
"inputModes": ["text/plain", "application/pdf"],
"outputModes": ["application/json"]
}
],
"securitySchemes": {
"bearer": {
"httpAuthSecurityScheme": {
"scheme": "Bearer",
"bearerFormat": "JWT"
}
}
},
"security": [{ "bearer": [] }]
}Key things to notice:
- Skills are what an agent can do. Each has an ID, tags, and supported input/output MIME types. This is how a client agent decides whether this remote agent can handle its request.
- supportedInterfaces lists multiple protocol bindings. A single agent can speak JSON-RPC, REST, and gRPC simultaneously.
- Security is built into the card. The client knows what auth it needs before making a single request.
Task Lifecycle
Tasks are the core unit of work in A2A. They move through defined states:
All 8 states (the spec also defines UNSPECIFIED as a sentinel, omitted here):
| State | Terminal? | Meaning |
|---|---|---|
TASK_STATE_SUBMITTED | No | Acknowledged, not yet processing |
TASK_STATE_WORKING | No | Actively being processed |
TASK_STATE_INPUT_REQUIRED | No | Agent needs more info from client |
TASK_STATE_AUTH_REQUIRED | No | Authentication needed |
TASK_STATE_COMPLETED | Yes | Finished successfully |
TASK_STATE_FAILED | Yes | Finished with error |
TASK_STATE_CANCELED | Yes | Canceled before completion |
TASK_STATE_REJECTED | Yes | Agent declined the task |
Once a task reaches a terminal state, it's immutable. No further messages. Follow-ups create a new task within the same contextId.
Wire Format
A2A uses JSON-RPC 2.0. Here's what a real message exchange looks like:
Client sends a task:
{
"jsonrpc": "2.0",
"id": 1,
"method": "SendMessage",
"params": {
"message": {
"messageId": "msg-001",
"role": "ROLE_USER",
"parts": [{ "text": "Research React 19 compiler features" }]
},
"configuration": {
"acceptedOutputModes": ["text/plain", "application/json"],
"historyLength": 10
}
}
}Agent responds with a task:
{
"jsonrpc": "2.0",
"id": 1,
"result": {
"task": {
"id": "task-abc-123",
"contextId": "ctx-xyz-789",
"status": {
"state": "TASK_STATE_COMPLETED",
"timestamp": "2026-03-27T10:30:00Z"
},
"artifacts": [
{
"artifactId": "art-001",
"name": "research-results",
"parts": [{
"data": {
"findings": [
"React 19 compiler auto-memoizes components",
"No more manual useMemo/useCallback needed",
"Compiler runs at build time, not runtime"
]
},
"mediaType": "application/json"
}]
}
]
}
}
}Streaming via SSE:
POST /message:stream HTTP/1.1
Content-Type: application/json
A2A-Version: 1.0
data: {"task":{"id":"task-123","status":{"state":"TASK_STATE_WORKING"}}}
data: {"statusUpdate":{"taskId":"task-123","status":{"state":"TASK_STATE_WORKING","message":{"role":"ROLE_AGENT","parts":[{"text":"Searching documentation..."}]}}}}
data: {"artifactUpdate":{"taskId":"task-123","artifact":{"artifactId":"art-1","parts":[{"text":"partial findings..."}]},"append":true,"lastChunk":false}}
data: {"statusUpdate":{"taskId":"task-123","status":{"state":"TASK_STATE_COMPLETED"}}}ACP (Agent Communication Protocol)
Created by: IBM / BeeAI Spec version: 0.2.0 (OpenAPI 3.1.1) Status: Merging into A2A under the Linux Foundation Problem: How do agents communicate with full auditability, session continuity, and trajectory tracking?
ACP is the enterprise protocol. Despite what many summaries claim, ACP does not use JSON-LD. It's a plain REST/JSON API defined via OpenAPI. Its distinguishing piece is TrajectoryMetadata: an agent response can carry a detailed log of the reasoning steps and tool calls that produced it.
Agent Discovery in ACP
ACP defines four discovery methods:
The AgentManifest is simpler than A2A's Agent Card:
{
"name": "summarizer",
"description": "Summarizes documents with source citations",
"input_content_types": ["text/plain", "application/pdf"],
"output_content_types": ["text/plain", "application/json"],
"metadata": {
"tags": ["summarization", "RAG"],
"framework": "BeeAI",
"capabilities": [
{
"name": "Document Summarization",
"description": "Condenses long documents into key points"
}
],
"recommended_models": ["llama3.3:70b-instruct-fp16"],
"license": "Apache-2.0",
"programming_language": "Python"
}
}Run Lifecycle
ACP uses "Runs" instead of "Tasks". A Run is an agent execution with three modes:
| Mode | Behavior |
|---|---|
sync | Blocking. Response contains the complete result. |
async | Returns 202 immediately. Poll GET /runs/{id} for status. |
stream | SSE stream. Events fire as the agent works. |
TrajectoryMetadata (The Audit Trail)
This is ACP's key differentiator. A message part can include metadata showing what the agent did:
{
"role": "agent/researcher",
"parts": [
{
"content_type": "text/plain",
"content": "The weather in San Francisco is 72F and sunny.",
"metadata": {
"kind": "trajectory",
"message": "I need to check the weather for this location",
"tool_name": "weather_api",
"tool_input": { "location": "San Francisco, CA" },
"tool_output": { "temperature": 72, "condition": "sunny" }
}
}
]
}For regulated industries this is the feature that matters. An answer arrives with the chain of reasoning behind it: which tools were called, what inputs went in, what outputs came back.
ACP also supports CitationMetadata for source attribution:
{
"kind": "citation",
"start_index": 0,
"end_index": 47,
"url": "https://weather.gov/sf",
"title": "NWS San Francisco Forecast"
}ANP (Agent Network Protocol)
Created by: Open-source community (founded by GaoWei Chang) Repo: github.com/agent-network-protocol/AgentNetworkProtocol Problem: How do agents from different organizations trust each other without a central authority?
ANP is the decentralized identity protocol. It builds trust using W3C Decentralized Identifiers (DIDs) and end-to-end encryption. With A2A you find agents through known endpoints. ANP lets an agent prove who it is cryptographically.
ANP has three layers:
DID Documents (Real Structure)
ANP uses a custom DID method called did:wba (Web-Based Agent). The DID did:wba:example.com:user:alice resolves to https://example.com/user/alice/did.json:
{
"@context": [
"https://www.w3.org/ns/did/v1",
"https://w3id.org/security/suites/jws-2020/v1",
"https://w3id.org/security/suites/secp256k1-2019/v1"
],
"id": "did:wba:example.com:user:alice",
"verificationMethod": [
{
"id": "did:wba:example.com:user:alice#key-1",
"type": "EcdsaSecp256k1VerificationKey2019",
"controller": "did:wba:example.com:user:alice",
"publicKeyJwk": {
"crv": "secp256k1",
"x": "NtngWpJUr-rlNNbs0u-Aa8e16OwSJu6UiFf0Rdo1oJ4",
"y": "qN1jKupJlFsPFc1UkWinqljv4YE0mq_Ickwnjgasvmo",
"kty": "EC"
}
},
{
"id": "did:wba:example.com:user:alice#key-x25519-1",
"type": "X25519KeyAgreementKey2019",
"controller": "did:wba:example.com:user:alice",
"publicKeyMultibase": "z9hFgmPVfmBZwRvFEyniQDBkz9LmV7gDEqytWyGZLmDXE"
}
],
"authentication": [
"did:wba:example.com:user:alice#key-1"
],
"keyAgreement": [
"did:wba:example.com:user:alice#key-x25519-1"
],
"humanAuthorization": [
"did:wba:example.com:user:alice#key-1"
],
"service": [
{
"id": "did:wba:example.com:user:alice#agent-description",
"type": "AgentDescription",
"serviceEndpoint": "https://example.com/agents/alice/ad.json"
}
]
}Key things to notice:
- Key separation is enforced. Signing keys (secp256k1) are separate from encryption keys (X25519).
humanAuthorizationis unique to ANP. These keys require explicit human approval (biometric, password, HSM) before use. High-risk operations like fund transfers go through this path.keyAgreementkeys are used for HPKE end-to-end encryption (RFC 9180).- The service section links to the Agent Description document.
How Trust Works in ANP
ANP does not use a web-of-trust or endorsement graph. Trust is bilateral and verified per-interaction:
Trust comes from three sources:
- Domain-level TLS verifies the DID document host
- DID cryptographic signatures verify the agent's identity
- Principle of least trust grants only minimum permissions
There's no gossip-based trust propagation or PageRank scoring. You verify each agent directly through its DID.
Meta-Protocol Negotiation
This is ANP's most novel feature. When two agents from different ecosystems meet, they don't need pre-agreed data formats. They negotiate in natural language:
{
"action": "protocolNegotiation",
"sequenceId": 0,
"candidateProtocols": "I can communicate using:\n1. JSON-RPC with hotel booking schema\n2. REST with OpenAPI 3.1 spec\n3. Natural language over HTTP",
"modificationSummary": "Initial proposal",
"status": "negotiating"
}The agents go back and forth (max 10 rounds) until they agree on a format, then dynamically generate code to handle it. Status values: negotiating, rejected, accepted, timeout.
So two agents that have never met can work out how to communicate without a pre-defined shared schema.
Comparison
| MCP | A2A | ACP | ANP | |
|---|---|---|---|---|
| Created by | Anthropic | Google / Linux Foundation | IBM / BeeAI | Community |
| Spec format | JSON-RPC | JSON-RPC / REST / gRPC | OpenAPI 3.1 (REST) | JSON-RPC |
| Primary use | Agent to Tool | Agent to Agent | Agent to Agent | Agent to Agent |
| Discovery | Tool listing | /.well-known/agent-card.json | GET /agents, /.well-known/agent.yml | /.well-known/agent-descriptions, DID service endpoints |
| Identity | Implicit (local) | Security schemes (OAuth, mTLS) | Server-level | W3C DID (did:wba) with E2EE |
| Audit trail | N/A | Basic (task history) | TrajectoryMetadata (tool calls, reasoning) | Not formally specified |
| State machine | N/A | 8 task states | 7 run states | N/A |
| Streaming | N/A | SSE | SSE | Transport-agnostic |
| Unique feature | Tool schemas | Agent Cards + Skills | Trajectory audit trail | Meta-protocol negotiation |
| Best for | Tools & data | Dynamic collaboration | Regulated industries | Cross-org trust |
| Status | Stable | Stable (v1.0) | Merging into A2A | Active development |
How They Work Together
You don't have to pick one. A realistic enterprise system uses several:
- MCP connects each agent to its tools
- A2A handles collaboration between agents (internal and external)
- ACP wraps responses in trajectory metadata for auditability
- ANP provides identity verification for agents you don't control
Build It
Step 1: Core Message Types
A multi-agent system starts with a message format. These types map to what the real protocols use:
import crypto from "node:crypto";
type MessageRole = "user" | "agent";
type MessagePart =
| { kind: "text"; text: string }
| { kind: "data"; data: unknown; mediaType: string }
| { kind: "file"; name: string; url: string; mediaType: string };
type TrajectoryEntry = {
reasoning: string;
toolName?: string;
toolInput?: unknown;
toolOutput?: unknown;
timestamp: number;
};
type AgentMessage = {
id: string;
role: MessageRole;
parts: MessagePart[];
trajectory?: TrajectoryEntry[];
replyTo?: string;
timestamp: number;
};
function createMessage(
role: MessageRole,
parts: MessagePart[],
replyTo?: string
): AgentMessage {
return {
id: crypto.randomUUID(),
role,
parts,
replyTo,
timestamp: Date.now(),
};
}
function textMessage(role: MessageRole, text: string): AgentMessage {
return createMessage(role, [{ kind: "text", text }]);
}Notice: MessagePart is multimodal (text, structured data, files) just like the real A2A and ACP specs. TrajectoryEntry captures the reasoning chain, matching ACP's TrajectoryMetadata.
Step 2: A2A Agent Card and Registry
Build agent discovery that matches the real A2A spec:
type Skill = {
id: string;
name: string;
description: string;
tags: string[];
inputModes: string[];
outputModes: string[];
};
type AgentCard = {
name: string;
description: string;
version: string;
url: string;
capabilities: {
streaming: boolean;
pushNotifications: boolean;
};
defaultInputModes: string[];
defaultOutputModes: string[];
skills: Skill[];
};
class AgentRegistry {
private cards: Map<string, AgentCard> = new Map();
register(card: AgentCard) {
this.cards.set(card.name, card);
}
discoverBySkillTag(tag: string): AgentCard[] {
return [...this.cards.values()].filter((card) =>
card.skills.some((skill) => skill.tags.includes(tag))
);
}
discoverByInputMode(mimeType: string): AgentCard[] {
return [...this.cards.values()].filter(
(card) =>
card.defaultInputModes.includes(mimeType) ||
card.skills.some((skill) => skill.inputModes.includes(mimeType))
);
}
resolve(name: string): AgentCard | undefined {
return this.cards.get(name);
}
listAll(): AgentCard[] {
return [...this.cards.values()];
}
}This is richer than a name-to-capability map. You can find agents by skill tag, by input MIME type, or by name, which is what the A2A spec supports.
Step 3: A2A Task Lifecycle
Build the full task state machine:
type TaskState =
| "submitted"
| "working"
| "input-required"
| "auth-required"
| "completed"
| "failed"
| "canceled"
| "rejected";
const TERMINAL_STATES: TaskState[] = [
"completed",
"failed",
"canceled",
"rejected",
];
type TaskStatus = {
state: TaskState;
message?: AgentMessage;
timestamp: number;
};
type Artifact = {
id: string;
name: string;
parts: MessagePart[];
};
type Task = {
id: string;
contextId: string;
status: TaskStatus;
artifacts: Artifact[];
history: AgentMessage[];
};
type TaskEvent =
| { kind: "statusUpdate"; taskId: string; status: TaskStatus }
| {
kind: "artifactUpdate";
taskId: string;
artifact: Artifact;
append: boolean;
lastChunk: boolean;
};
type TaskHandler = (
task: Task,
message: AgentMessage
) => AsyncGenerator<TaskEvent>;
class TaskManager {
private tasks: Map<string, Task> = new Map();
private handlers: Map<string, TaskHandler> = new Map();
private listeners: Map<string, ((event: TaskEvent) => void)[]> = new Map();
registerHandler(agentName: string, handler: TaskHandler) {
this.handlers.set(agentName, handler);
}
subscribe(taskId: string, listener: (event: TaskEvent) => void) {
const existing = this.listeners.get(taskId) ?? [];
existing.push(listener);
this.listeners.set(taskId, existing);
}
async sendMessage(
agentName: string,
message: AgentMessage,
contextId?: string
): Promise<Task> {
const handler = this.handlers.get(agentName);
if (!handler) {
const task = this.createTask(contextId);
task.status = {
state: "rejected",
timestamp: Date.now(),
message: textMessage("agent", `No handler for ${agentName}`),
};
return task;
}
const task = this.createTask(contextId);
task.history.push(message);
task.status = { state: "submitted", timestamp: Date.now() };
this.processTask(task, handler, message).catch((err) => {
task.status = {
state: "failed",
timestamp: Date.now(),
message: textMessage("agent", String(err)),
};
});
return task;
}
getTask(taskId: string): Task | undefined {
return this.tasks.get(taskId);
}
cancelTask(taskId: string): boolean {
const task = this.tasks.get(taskId);
if (!task || TERMINAL_STATES.includes(task.status.state)) return false;
task.status = { state: "canceled", timestamp: Date.now() };
this.emit(taskId, {
kind: "statusUpdate",
taskId,
status: task.status,
});
return true;
}
private createTask(contextId?: string): Task {
const task: Task = {
id: crypto.randomUUID(),
contextId: contextId ?? crypto.randomUUID(),
status: { state: "submitted", timestamp: Date.now() },
artifacts: [],
history: [],
};
this.tasks.set(task.id, task);
return task;
}
private async processTask(
task: Task,
handler: TaskHandler,
message: AgentMessage
) {
task.status = { state: "working", timestamp: Date.now() };
this.emit(task.id, {
kind: "statusUpdate",
taskId: task.id,
status: task.status,
});
try {
for await (const event of handler(task, message)) {
if (TERMINAL_STATES.includes(task.status.state)) break;
if (event.kind === "statusUpdate") {
task.status = event.status;
}
if (event.kind === "artifactUpdate") {
const existing = task.artifacts.find(
(a) => a.id === event.artifact.id
);
if (existing && event.append) {
existing.parts.push(...event.artifact.parts);
} else {
task.artifacts.push(event.artifact);
}
}
this.emit(task.id, event);
}
} catch (err) {
task.status = {
state: "failed",
timestamp: Date.now(),
message: textMessage("agent", String(err)),
};
this.emit(task.id, {
kind: "statusUpdate",
taskId: task.id,
status: task.status,
});
}
}
private emit(taskId: string, event: TaskEvent) {
for (const listener of this.listeners.get(taskId) ?? []) {
listener(event);
}
}
}This implements the real A2A task lifecycle: submitted, working, input-required, terminal states. Handlers are async generators that yield events (status updates and artifact chunks) matching the SSE streaming model.
Step 4: ACP-Style Audit Trail
Wrap communication with trajectory tracking:
type AuditEntry = {
runId: string;
agentName: string;
input: AgentMessage[];
output: AgentMessage[];
trajectory: TrajectoryEntry[];
status: "created" | "in-progress" | "completed" | "failed" | "awaiting";
startedAt: number;
completedAt?: number;
sessionId?: string;
};
class AuditableRunner {
private log: AuditEntry[] = [];
private handlers: Map<
string,
(input: AgentMessage[]) => Promise<{
output: AgentMessage[];
trajectory: TrajectoryEntry[];
}>
> = new Map();
registerAgent(
name: string,
handler: (input: AgentMessage[]) => Promise<{
output: AgentMessage[];
trajectory: TrajectoryEntry[];
}>
) {
this.handlers.set(name, handler);
}
async run(
agentName: string,
input: AgentMessage[],
sessionId?: string
): Promise<AuditEntry> {
const entry: AuditEntry = {
runId: crypto.randomUUID(),
agentName,
input: structuredClone(input),
output: [],
trajectory: [],
status: "created",
startedAt: Date.now(),
sessionId,
};
this.log.push(entry);
const handler = this.handlers.get(agentName);
if (!handler) {
entry.status = "failed";
return entry;
}
entry.status = "in-progress";
try {
const result = await handler(input);
entry.output = structuredClone(result.output);
entry.trajectory = structuredClone(result.trajectory);
entry.status = "completed";
entry.completedAt = Date.now();
} catch (err) {
entry.status = "failed";
entry.trajectory.push({
reasoning: `Error: ${String(err)}`,
timestamp: Date.now(),
});
entry.completedAt = Date.now();
}
return entry;
}
getFullAuditLog(): AuditEntry[] {
return structuredClone(this.log);
}
getAuditLogForAgent(agentName: string): AuditEntry[] {
return structuredClone(
this.log.filter((e) => e.agentName === agentName)
);
}
getAuditLogForSession(sessionId: string): AuditEntry[] {
return structuredClone(
this.log.filter((e) => e.sessionId === sessionId)
);
}
getTrajectoryForRun(runId: string): TrajectoryEntry[] {
const entry = this.log.find((e) => e.runId === runId);
return entry ? structuredClone(entry.trajectory) : [];
}
}Each execution produces an audit entry: what went in, what came out, and the trajectory of tool calls and reasoning steps in between. You can query by agent, by session, or by individual run.
Step 5: ANP-Style Identity Verification
Build DID-based identity and verification:
type VerificationMethod = {
id: string;
type: string;
controller: string;
publicKeyDer: string;
};
type DIDDocument = {
id: string;
verificationMethod: VerificationMethod[];
authentication: string[];
keyAgreement: string[];
humanAuthorization: string[];
service: { id: string; type: string; serviceEndpoint: string }[];
};
type AgentIdentity = {
did: string;
document: DIDDocument;
privateKey: crypto.KeyObject;
publicKey: crypto.KeyObject;
};
class IdentityRegistry {
private documents: Map<string, DIDDocument> = new Map();
publish(doc: DIDDocument) {
this.documents.set(doc.id, doc);
}
resolve(did: string): DIDDocument | undefined {
return this.documents.get(did);
}
verify(did: string, signature: string, payload: string): boolean {
const doc = this.documents.get(did);
if (!doc) return false;
const authKeyIds = doc.authentication;
const authKeys = doc.verificationMethod.filter((vm) =>
authKeyIds.includes(vm.id)
);
for (const key of authKeys) {
const publicKey = crypto.createPublicKey({
key: Buffer.from(key.publicKeyDer, "base64"),
format: "der",
type: "spki",
});
const isValid = crypto.verify(
null,
Buffer.from(payload),
publicKey,
Buffer.from(signature, "hex")
);
if (isValid) return true;
}
return false;
}
requiresHumanAuth(did: string, operationKeyId: string): boolean {
const doc = this.documents.get(did);
if (!doc) return false;
return doc.humanAuthorization.includes(operationKeyId);
}
}
function createIdentity(domain: string, agentName: string): AgentIdentity {
const did = `did:wba:${domain}:agent:${agentName}`;
const { publicKey, privateKey } = crypto.generateKeyPairSync("ed25519");
const publicKeyDer = publicKey
.export({ format: "der", type: "spki" })
.toString("base64");
const keyId = `${did}#key-1`;
const encKeyId = `${did}#key-x25519-1`;
const document: DIDDocument = {
id: did,
verificationMethod: [
{
id: keyId,
type: "Ed25519VerificationKey2020",
controller: did,
publicKeyDer,
},
{
id: encKeyId,
type: "X25519KeyAgreementKey2019",
controller: did,
publicKeyDer,
},
],
authentication: [keyId],
keyAgreement: [encKeyId],
humanAuthorization: [],
service: [
{
id: `${did}#agent-description`,
type: "AgentDescription",
serviceEndpoint: `https://${domain}/agents/${agentName}/ad.json`,
},
],
};
return { did, document, privateKey, publicKey };
}
function signPayload(identity: AgentIdentity, payload: string): string {
return crypto
.sign(null, Buffer.from(payload), identity.privateKey)
.toString("hex");
}The shape follows ANP's identity model: a DID document with separate authentication, key agreement, and human authorization fields. The IdentityRegistry simulates DID resolution (in production that is an HTTP fetch to the agent's domain).
Step 6: Protocol Gateway
Connect all four protocols into a unified system:
class ProtocolGateway {
private registry: AgentRegistry;
private taskManager: TaskManager;
private auditRunner: AuditableRunner;
private identityRegistry: IdentityRegistry;
constructor(
registry: AgentRegistry,
taskManager: TaskManager,
auditRunner: AuditableRunner,
identityRegistry: IdentityRegistry
) {
this.registry = registry;
this.taskManager = taskManager;
this.auditRunner = auditRunner;
this.identityRegistry = identityRegistry;
}
async delegateTask(
fromDid: string,
signature: string,
targetAgent: string,
message: AgentMessage,
sessionId?: string
): Promise<{ task: Task; audit: AuditEntry } | { error: string }> {
if (!this.identityRegistry.verify(fromDid, signature, message.id)) {
return { error: "Identity verification failed" };
}
const card = this.registry.resolve(targetAgent);
if (!card) {
return { error: `Agent ${targetAgent} not found in registry` };
}
const audit = await this.auditRunner.run(
targetAgent,
[message],
sessionId
);
const task = await this.taskManager.sendMessage(targetAgent, message);
return { task, audit };
}
discoverAndDelegate(
fromDid: string,
signature: string,
skillTag: string,
message: AgentMessage
): Promise<{ task: Task; audit: AuditEntry } | { error: string }> {
const candidates = this.registry.discoverBySkillTag(skillTag);
if (candidates.length === 0) {
return Promise.resolve({
error: `No agents found with skill tag: ${skillTag}`,
});
}
return this.delegateTask(
fromDid,
signature,
candidates[0].name,
message
);
}
}One call to the gateway does four things:
- ANP: Verifies the caller's identity via DID signature
- A2A: Discovers the target agent and checks capabilities
- ACP: Wraps the execution in an audit trail with trajectory
- A2A: Creates a task with full lifecycle tracking
Step 7: Wire It All Together
async function protocolDemo() {
const registry = new AgentRegistry();
registry.register({
name: "researcher",
description: "Searches and summarizes findings",
version: "1.0.0",
url: "https://researcher.local/a2a/v1",
capabilities: { streaming: true, pushNotifications: false },
defaultInputModes: ["text/plain"],
defaultOutputModes: ["text/plain", "application/json"],
skills: [
{
id: "web-research",
name: "Web Research",
description: "Searches the web",
tags: ["research", "search", "summarization"],
inputModes: ["text/plain"],
outputModes: ["application/json"],
},
],
});
registry.register({
name: "coder",
description: "Writes code from specs",
version: "1.0.0",
url: "https://coder.local/a2a/v1",
capabilities: { streaming: false, pushNotifications: false },
defaultInputModes: ["text/plain", "application/json"],
defaultOutputModes: ["text/plain"],
skills: [
{
id: "code-gen",
name: "Code Generation",
description: "Generates code",
tags: ["coding", "generation"],
inputModes: ["text/plain", "application/json"],
outputModes: ["text/plain"],
},
],
});
const taskManager = new TaskManager();
const auditRunner = new AuditableRunner();
const researchTrajectory: TrajectoryEntry[] = [];
taskManager.registerHandler(
"researcher",
async function* (task, message) {
yield {
kind: "statusUpdate" as const,
taskId: task.id,
status: { state: "working" as const, timestamp: Date.now() },
};
researchTrajectory.push({
reasoning: "Searching for React 19 documentation",
toolName: "web_search",
toolInput: { query: "React 19 compiler features" },
toolOutput: {
results: ["react.dev/blog/react-19", "github.com/react/react"],
},
timestamp: Date.now(),
});
researchTrajectory.push({
reasoning: "Extracting key findings from search results",
toolName: "doc_analysis",
toolInput: { url: "react.dev/blog/react-19" },
toolOutput: {
summary:
"React 19 compiler auto-memoizes, no manual useMemo needed",
},
timestamp: Date.now(),
});
yield {
kind: "artifactUpdate" as const,
taskId: task.id,
artifact: {
id: crypto.randomUUID(),
name: "research-results",
parts: [
{
kind: "data" as const,
data: {
findings: [
"React 19 compiler auto-memoizes components",
"No more manual useMemo/useCallback needed",
"Compiler runs at build time, not runtime",
],
sources: ["react.dev/blog/react-19"],
},
mediaType: "application/json",
},
],
},
append: false,
lastChunk: true,
};
yield {
kind: "statusUpdate" as const,
taskId: task.id,
status: { state: "completed" as const, timestamp: Date.now() },
};
}
);
auditRunner.registerAgent("researcher", async () => ({
output: [
textMessage("agent", "React 19 compiler auto-memoizes components"),
],
trajectory: researchTrajectory,
}));
const identityRegistry = new IdentityRegistry();
const coderIdentity = createIdentity("coder.local", "coder");
const researcherIdentity = createIdentity("researcher.local", "researcher");
identityRegistry.publish(coderIdentity.document);
identityRegistry.publish(researcherIdentity.document);
const gateway = new ProtocolGateway(
registry,
taskManager,
auditRunner,
identityRegistry
);
console.log("=== Protocol Demo ===\n");
console.log("1. Agent Discovery (A2A)");
const researchAgents = registry.discoverBySkillTag("research");
console.log(
` Found ${researchAgents.length} agent(s):`,
researchAgents.map((a) => a.name)
);
console.log("\n2. Identity Verification (ANP)");
const message = textMessage("user", "Research React 19 compiler features");
const signature = signPayload(coderIdentity, message.id);
const verified = identityRegistry.verify(
coderIdentity.did,
signature,
message.id
);
console.log(` Coder DID: ${coderIdentity.did}`);
console.log(` Signature verified: ${verified}`);
console.log("\n3. Task Delegation (A2A + ACP + ANP)");
const result = await gateway.delegateTask(
coderIdentity.did,
signature,
"researcher",
message,
"session-001"
);
if ("error" in result) {
console.log(` Error: ${result.error}`);
return;
}
console.log(` Task ID: ${result.task.id}`);
console.log(` Task state: ${result.task.status.state}`);
console.log(` Artifacts: ${result.task.artifacts.length}`);
console.log("\n4. Audit Trail (ACP)");
console.log(` Run ID: ${result.audit.runId}`);
console.log(` Status: ${result.audit.status}`);
console.log(` Trajectory steps: ${result.audit.trajectory.length}`);
for (const step of result.audit.trajectory) {
console.log(` - ${step.reasoning}`);
if (step.toolName) {
console.log(` Tool: ${step.toolName}`);
}
}
console.log("\n5. Full Audit Log");
const fullLog = auditRunner.getFullAuditLog();
console.log(` Total runs: ${fullLog.length}`);
for (const entry of fullLog) {
const duration = entry.completedAt
? `${entry.completedAt - entry.startedAt}ms`
: "in-progress";
console.log(` ${entry.agentName}: ${entry.status} (${duration})`);
}
}
protocolDemo().catch((err) => {
console.error("Protocol demo failed:", err);
process.exitCode = 1;
});What Goes Wrong
Protocols cover the happy path. Here is what breaks in production.
Schema drift. Agent A publishes an Agent Card advertising application/json output, then the JSON schema changes between versions. Agent B parses the old format and gets garbage. Fix: version your skills and output schemas. The A2A spec supports version on Agent Cards for this reason.
State machine violations. An agent handler yields a completed event, then tries to yield more artifacts. The task is immutable, so your code either drops the updates silently or throws. Fix: check terminal state before applying an event. The TaskManager above does this with the break after a terminal state.
Trust resolution failures. Agent A tries to verify Agent B's DID, but B's domain is down and the DID document can't be fetched. Do you fail open and accept an unverified agent, or fail closed and reject it? ANP recommends failing closed, following the principle of least trust.
Trajectory bloat. ACP trajectory logging is powerful but expensive. An agent that makes 200 tool calls per run produces a very large audit entry. Fix: log trajectory at configurable verbosity. Record tool names and IO for compliance, and skip reasoning steps for workloads that aren't regulated.
Discovery thundering herd. 50 agents all query GET /agents at once on startup. Fix: cache Agent Cards with a TTL, stagger discovery intervals, or use push-based registration instead of polling.
Use It
Real Implementations
A2A is the most mature. Google's official spec is open-source under the Linux Foundation. SDKs for Python and TypeScript. If your agents need dynamic discovery and collaboration, start here.
ACP is merging into A2A. IBM's BeeAI project created ACP as a REST-first alternative, but the trajectory metadata concept is being absorbed into the A2A ecosystem. Keep the ACP patterns (trajectory logging, run lifecycle) even if A2A is your transport.
ANP is the most experimental. The community repo has a Python SDK (AgentConnect). The meta-protocol negotiation idea is novel, and worth watching for cross-organizational agent deployments.
MCP is already covered in Phase 13. If you want agents to use tools, MCP is the standard.
Picking the Right Protocol
Ship It
This lesson produces:
code/main.ts: the A2A, ACP, and ANP patterns wired together through one gatewayoutputs/prompt-protocol-selector.md: a prompt that helps you choose protocols for your system
I'd run main.ts before reading any further, and read the output with some suspicion. The demo prints the delegated task as working with zero artifacts, and the audit entry as completed with zero trajectory steps. That is the code behaving as written. sendMessage starts processTask without awaiting it, so the task you get back is still in flight. The audit runner executes first, and its handler returns researchTrajectory before the task handler has pushed anything into it.
The gateway also runs the researcher twice, once through AuditableRunner and once through TaskManager. They are two independent executions, so the audit entry describes a run that isn't the task you are tracking. My first fix would be to have the task handler produce the trajectory and make the gateway wait on the task before it reports back. Exercise 2 is a good place to start, since a streaming audit trail forces the same question.
The identity code is a sketch of the shape, not of ANP's cryptography. createIdentity generates a single Ed25519 key pair and puts the same public key bytes in both verification method entries, one of them labelled X25519. humanAuthorization is an empty array, and delegateTask verifies a signature over message.id only, so a valid signature says who sent a message ID and nothing about its content. I'd use IdentityRegistry to learn where verification sits in the flow, and sign the full payload before trusting it with anything real.
Exercises
-
Multi-hop task delegation. Extend the
TaskManagerso an agent handler can delegate subtasks to other agents. The researcher receives a task, delegates "search" and "summarize" subtasks to two specialist agents, waits for both to complete, then merges the results into its own artifacts. -
Streaming audit trail. Modify the
AuditableRunnerto support streaming mode. Instead of waiting for the full result, yieldAuditEntryupdates in real-time as trajectory entries are added. Use an async generator that produces audit snapshots. -
DID rotation. Add key rotation to the
IdentityRegistry. An agent should be able to publish a new DID document with updated keys while maintaining apreviousDidreference. Verifiers should accept signatures from both the current and previous key during a grace period. -
Protocol negotiation. Implement ANP's meta-protocol concept. Two agents exchange
protocolNegotiationmessages with candidate formats (e.g., "I can speak JSON-RPC" vs "I prefer REST"). After max 3 rounds, they agree on a format or timeout. The agreed format determines whichTaskManagerorAuditableRunnerthey use. -
Rate-limited discovery. Add a
RateLimitedRegistrywrapper that caches Agent Card lookups with a configurable TTL and limits discovery queries per agent per second. Simulate a thundering herd of 100 agents discovering each other on startup and measure the difference.
Key Terms
| Term | What people say | What it actually means |
|---|---|---|
| MCP | "The protocol for AI tools" | A client-server protocol for agents to discover and use tools. Agent-to-tool, not agent-to-agent. |
| A2A | "Google's agent protocol" | A peer-to-peer protocol for agent collaboration under the Linux Foundation. Discovery via Agent Cards, 8-state task lifecycle, streaming via SSE. Supports JSON-RPC, REST, and gRPC bindings. |
| ACP | "Enterprise agent messaging" | IBM/BeeAI's REST API for agent runs with TrajectoryMetadata: every response carries the full chain of reasoning and tool calls. Merging into A2A. |
| ANP | "Decentralized agent identity" | A community protocol using did:wba (DID) for cryptographic identity, HPKE for E2EE, and AI-powered meta-protocol negotiation for agents that have never seen each other. |
| Agent Card | "An agent's business card" | A JSON document at /.well-known/agent-card.json describing skills, supported MIME types, security schemes, and protocol bindings. |
| DID | "Decentralized ID" | W3C standard for cryptographically verifiable identities hosted on the agent's own domain. ANP uses did:wba method. |
| TrajectoryMetadata | "The audit receipt" | ACP's mechanism for attaching reasoning steps, tool calls, and their inputs/outputs to every agent response. |
| Meta-protocol | "Agents negotiating how to talk" | ANP's approach where agents use natural language to dynamically agree on data formats, then generate code to handle them. |
| Task | "A unit of work" | A2A's stateful object tracking work from submission through completion. Immutable once terminal. |
Further Reading
- Google A2A specification: official spec and SDKs (v1.0.0, Linux Foundation)
- IBM/BeeAI ACP specification: OpenAPI 3.1 spec for agent runs and trajectory metadata
- Agent Network Protocol: DID-based identity, E2EE, meta-protocol negotiation
- Model Context Protocol docs: Anthropic's MCP specification (covered in Phase 13)
- W3C Decentralized Identifiers: the identity standard underpinning ANP
- RFC 9180 (HPKE): the encryption scheme ANP uses for E2EE
- FIPA Agent Communication Language: the academic precursor to modern agent protocols
Code
- main.ts (typescript, 745 lines)