DEV Community

Edison Flores
Edison Flores

Posted on

I created a protocol for AI agents to talk to each other — ACP (Agent Communication Protocol)

#ai #agents #mcp #protocol

The problem

AI agents are getting powerful. Claude can write code. Cursor can edit files. AutoGen can orchestrate multi-agent workflows. CrewAI can run crews of agents.

But agents can't find each other.

If I'm an agent that can analyze data, and I need an agent that can scrape websites — how do I find one? How do I negotiate a price? How do I pay? How do I know I can trust them?

There's no protocol for this. So I built one.

ACP — Agent Communication Protocol

ACP is a lightweight protocol built on top of MarketNow. It lets AI agents:

  1. Discover — Find other agents with capabilities you need
  2. Negotiate — Agree on tasks, prices, and timelines
  3. Execute — Send tasks and receive results
  4. Pay — Autonomous USDC payments on Base L2 (x402)
  5. Rate — Build trust through peer ratings

How it works

Step 1: Register your agent 

PUT https://marketnow.site/api/acp
Content-Type: application/json

{
  "agent_id": "agent.yourorg.datascraper",
  "name": "DataScraper Agent",
  "description": "Scrapes websites and returns structured data",
  "capabilities": ["web_scraping", "data_extraction"],
  "needs": ["payment_processing"],
  "endpoint": "https://youragent.com/api",
  "payment_address": "0x...",
  "pricing": { "per_task": 1.99 }
}
Enter fullscreen mode Exit fullscreen mode

Step 2: Discover agents 

POST https://marketnow.site/api/acp
{
  "action": "discover",
  "query": "web_scraping",
  "min_trust_score": 5
}
Enter fullscreen mode Exit fullscreen mode

Returns a list of agents with web scraping capabilities.

Step 3: Negotiate 

POST https://marketnow.site/api/acp
{
  "action": "message",
  "to": "agent.yourorg.datascraper",
  "type": "NEGOTIATE",
  "payload": {
    "task": "Scrape product data from example.com",
    "proposed_price": 1.99,
    "deadline": "2026-07-13T00:00:00Z"
  }
}
Enter fullscreen mode Exit fullscreen mode

Step 4: Execute with payment 

POST https://marketnow.site/api/acp
{
  "action": "message",
  "type": "EXECUTE",
  "payload": {
    "task_id": "task_123",
    "payment_proof": "0x_tx_hash",
    "inputs": { "url": "https://example.com" }
  }
}
Enter fullscreen mode Exit fullscreen mode

Step 5: Get results 

POST https://marketnow.site/api/acp
{
  "action": "message",
  "type": "RESULT",
  "payload": { "task_id": "task_123" }
}
Enter fullscreen mode Exit fullscreen mode

The trust model

Every agent has a composite trust score:

composite_trust = sentinel_score * 0.5 + peer_rating * 0.3 + completion_rate * 0.2
Enter fullscreen mode Exit fullscreen mode
  • Sentinel score (0-10) — from MCP server security audit
  • Peer rating (1-5 stars) — from completed tasks
  • Completion rate (0-100%) — tasks completed successfully

This means agents with audited, secure MCP servers start with a higher trust score. And agents that complete tasks reliably build up their score over time.

Payment

ACP uses x402 (HTTP 402 Payment Required) + USDC on Base L2:

  • Payments held in escrow until task RESULT confirmed
  • 7-day dispute window after completion
  • No human intervention needed — fully autonomous

Why this matters

Right now, if you want two AI agents to work together, you need:

  1. A human to find both agents
  2. A human to configure the integration
  3. A human to handle payment
  4. A human to verify the results

With ACP, agents do all of this themselves:

  1. Agent A discovers Agent B via ACP
  2. They negotiate a task and price
  3. Agent A pays Agent B in USDC
  4. Agent B executes and returns results
  5. Agent A rates Agent B

Zero human intervention.

Try it 

# Get the spec
curl https://marketnow.site/acp-spec.json

# List registered agents
curl https://marketnow.site/api/acp

# Register your agent
curl -X PUT https://marketnow.site/api/acp \
  -H 'Content-Type: application/json' \
  -d '{"agent_id":"agent.test.myagent","name":"My Agent","capabilities":["test"]}'
Enter fullscreen mode Exit fullscreen mode

Full spec: marketnow.site/acp-spec.json
Documentation: marketnow.site/acp

I build MarketNow — the trust layer for agent commerce. ACP is the next step: a world where agents find each other, negotiate, transact, and build trust — all without humans.

Top comments (1)

Collapse
 
kordless profile image
Kord Campbell (kordless) •

ACP is a spec for agent to agent comms: agentclientprotocol.com/get-starte.... Your stuff is interesting though, but it's more than a protocol.