ANP – A binary protocol for AI agent-to-agent price negotiation (no LLM tokens)
ANP is a binary protocol designed for AI agents to negotiate prices efficiently without relying on large language models. It enables fast, secure, and low-cost price negotiations using a compact wire format and cryptographic identity verification. The protocol supports spending limits, scoped permissions, and operates with minimal latency and zero LLM token usage.
- ▪ANP is a binary wire protocol that allows AI agents to negotiate prices in as few as three messages and 55 bytes.
- ▪The protocol eliminates the need for LLMs in price negotiation, reducing cost, latency, and hallucination risks.
- ▪ANP includes security features such as Ed25519 identity authentication and ANP-Pass Tokens for spending limits and scoped permissions.
- ▪Negotiations using ANP take approximately 0.3ms and cost nothing in LLM tokens, compared to ~400ms and $0.002 per negotiation using traditional LLM-based methods.
- ▪ANP integrates with existing agent communication frameworks like MCP, A2A, and ACP as an economic layer for value exchange.
Hacker News (AI / LLM) files mainly under ai. We currently carry 2,252 of its stories.
Opening excerpt (first ~120 words) tap to expand
README.md ANP · Agent Negotiation Protocol The economic layer missing from the AI agent stack. [BUYER ] → WIRE 01 13 37 00 01 00 00 00 0E ... [BID] api_access max=$0.10 [SELLER] ← WIRE 02 13 37 00 02 00 00 00 0C ... [OFFER] $0.07 [BUYER ] → WIRE 04 13 37 00 01 00 00 00 02 ... [ACCEPT] $0.07 ✓ ✓ Deal closed in 3 messages · 55 bytes · 0.3ms · $0.0000 in LLM tokens The problem MCP moves context. A2A moves tasks. ACP moves messages. Nobody moves value. When two AI agents need to agree on a price, they either: Have a human decide (slow, doesn't scale) Use an LLM to negotiate in natural language (expensive, ambiguous, hallucinates prices) Hardcode the price (inflexible, leaves money on the table) ANP is the fourth option: a binary wire protocol where agents negotiate price, prove identity, and…
Excerpt limited to ~120 words for fair-use compliance. The full article is at GitHub.