The missing layer — v0.2.0 · Apache 2.0

AI agents can talk. They can pay. They can't yet negotiate.

In the next 24 months, a growing share of B2B procurement, renewals, and supplier negotiations will run agent-to-agent. Pactum already negotiates for Walmart and Maersk. But when two agents from different companies meet, there is no neutral protocol for them to commit on terms — no shared offer format, no way to verify the other side has authority, no tamper-evident record of what was agreed. A2CN is that layer.

See it in action → Read the spec
$20M
Pactum's Series A for agent procurement negotiation — a closed system serving Walmart and Maersk
3
Adjacent protocols exist: MCP, A2A, AP2. None covers commercial negotiation across organizational boundaries.
0
Open, neutral standards for how agents safely commit to binding commercial terms today.
Explore the protocol

Three ways to understand what A2CN does.

Scrub the timeline to see why now is the right moment. Run a live negotiation between two agents. Or wire it into a real procurement platform and see the adapter output change.

The negotiation layer needs to exist now.
Agent protocols have been converging on this gap for three years. A2CN fills it.
2022
Pactum raises $20M Series A
An AI agent negotiates tail-spend contracts for Walmart and Maersk. Proof the market exists. Closed platform; buyer-side only.
2024
Anthropic ships MCP
Agent-to-tool communication becomes a standard. Agents can now reliably call APIs and use structured tools — the plumbing layer.
2025 · spring
Google ships A2A
Agent-to-agent communication becomes a standard. Agents can discover each other and delegate tasks. But still no commercial layer.
2025 · fall
AP2 and ACP emerge
Agent payment execution gets standardized. Now agents can move money. The stack has talk, discover, pay — and a hole where negotiation should be.
2026
A2CN v0.2.0 — you are here
Open, neutral protocol for commercial negotiation between agents from different organizations. 8 components, 202 tests, Fairmarkit + Keelvar + Salesforce adapters, MCP server, A2A extension proposal submitted.
next
Meeting Place — neutral transaction hosting
When a design partner needs it: a hosted layer where agents from different orgs can run A2CN sessions without either side self-hosting. Currently planned, not built.
Watch a bilateral SaaS renewal negotiation.
Two agents. Different organizations. Neither controls the authoritative record — the cryptography does.
Ready
T
TechCorp Buyer Agent
did:web:techcorp.com
✓ Mandate verified
A
Acme Seller Agent
did:web:acme.io
✓ Mandate verified
Dual-signed transaction record generated independently by both parties
final terms SaaS renewal · $105,000 · net-45
buyer hash —
seller hash —
match ✓ identical — cryptographically binding on both sides
Drop A2CN into a procurement platform.
Zero platform changes. Adapters translate webhook events into A2CN sessions and agreed terms back into response format.
$1.80M
14 days
adapters/fairmarkit_adapter.py event: BID_CREATED
Where A2CN fits

The agent stack has three layers. A2CN is the fourth.

MCP — Agent-to-tool ✓ Exists Model Context Protocol — tools, resources, prompts A2A — Agent-to-agent communication ✓ Exists Google Agent-to-Agent — task delegation, capability discovery A2CN — Commercial negotiation ← this layer Offer exchange · Mandate verification · Dual-signed record AP2 / ACP — Payment execution ✓ Exists Agentic Payment Protocol / Agent Commerce Protocol — settlement

Agents can talk. Agents can pay. Agents cannot yet safely negotiate commercial terms across organizational boundaries.

Eight protocol components

What the spec actually covers.

Each component solves a specific problem that shows up the moment you try to wire two agents from different companies together to commit on a deal.

01
Discovery
/.well-known/a2cn-agent — agents advertise capabilities and find each other without a central registry
02
Mandate verification
Cryptographic proof an agent has authority to commit its organization. W3C DIDs; two-tier: Declared + Verifiable Credential.
03
Session invitation
Push-based handshake enabling adoption by parties without pre-deployed endpoints — cold-start solved
04
Offer exchange
Signed offers and counteroffers with strict turn-taking. ES256 + RFC 8785 JCS canonicalization.
05
Deal-type terms
Normative schemas for goods_procurement and saas_renewal, with extensible custom_terms for anything else
06
Session state machine
Turn enforcement, sequence ordering, round limits, impasse detection — formally specified transitions
07
Transaction record
Dual-signed, content-addressed record independently generated by both parties. Neither controls the authoritative copy.
08
Audit log
Structured EU AI Act compliance output for every terminal session state — COMPLETED, REJECTED, WITHDRAWN, IMPASSE
Where it stands

Built. Being proposed. Still to come.

A2CN is a solo build by one person with GTM experience across 5+ AI/ML startups — built from the sales seat, not the engineering org. The reference implementation has 202 passing tests. The A2A extension proposal is in review.

Shipped
7
  • Protocol spec v0.2.0 3,700+ lines, 8 components, Apache 2.0
  • Reference implementation Python, 202 tests passing
  • Session invitation cold-start adoption solved
  • Platform adapters Fairmarkit, Keelvar, Salesforce Revenue Cloud
  • MCP server Claude Desktop, LangChain, LangGraph, Agentforce, Cursor
  • JWT request authentication ES256, anti-replay, DID verification
  • LLM integration pattern Section 13.9, separation of concerns
In progress
1
  • A2A extension proposal OQ-011, submitted — borrows A2A's distribution
Planned
3
  • Meeting Place neutral transaction hosting, v0.3 — built when a design partner needs it
  • TypeScript reference implementation
  • SDK pip + npm
Who's building this
Christian
A2CN · solo author · Austin, TX
Early GTM at 5+ AI/ML startups — including Samsung Next and Comet ML
Spent years watching enterprise buyers evaluate AI tooling from the sales seat — what actually blocks deployment
Built A2CN's spec, reference implementation, platform adapters, and MCP server solo

I'm not a protocol author by training — I'm a GTM person who spent years selling AI tooling into enterprise procurement. The more I thought about where this is all going, the more it kept coming back to the same question: the agents can do the work, but there have to be guardrails for how they safely commit across company lines.

A2CN is my attempt to actually build what that future needs. The spec is rigorous because it has to be — signing, DIDs, dual-signed records, formally specified state machines. But the reason it exists isn't academic. It's because the agents are going to show up, and the guardrails need to be there before they do.

If you're a PM at a procurement platform, a contract AI company, or an agent framework wondering how to handle cross-organizational commitment — I want to talk. That's the bottleneck right now, not code.

Partner with us

If your platform is about to hit this problem, talk to us.

Design partner conversations are the active bottleneck. If you're running a procurement platform, a contract AI, or an agent framework that will soon need to commit on terms across organizational boundaries, we want to hear from you.

Read the spec
3,700+ lines covering eight protocol components, normative JSON schemas, platform integration patterns, and a complete bilateral walkthrough.
spec/a2cn-spec-v0.2.0.md →
Build on it
Python reference implementation with 202 passing tests. LangChain, CrewAI, and Agentforce compatible. Apache 2.0 license.
github.com/A2CN-protocol/A2CN →
Become a design partner
Pilot A2CN with your platform. Shape the v0.3 spec. Get integration support directly from the maintainer.
contact@a2cn.io →
Use with an LLM agent
Reference negotiation skills file for LLM-based A2CN agents. Grounded in MIT/Hopkins research on 182K agent-to-agent negotiations. Drop-in for Claude Code, LangChain, CrewAI.
reference-implementation/skills/ →