jiayaoqijia Agent Negotiation Protocol
Quick Reference
| Aspect | Detail |
|---|---|
| Pattern | StateGraph multi-round cyclic workflow |
| Parties | Buyer agent ↔ Seller agent |
| Strategy | Progressive concession (game-theoretic ZOPA convergence) |
| Settlement | x402 payment protocol (HTTP 402 → sign → verify → pay) |
| Reputation | ERC-8004 identity scoring (exponential moving average) |
| Max Rounds | 5 (configurable) |
| Scripts | scripts/negotiation_engine.py, scripts/x402_settlement.py |
| Examples | examples/basic_negotiation.py, examples/custom_strategy.py |
Decision Framework
When implementing agent negotiation, follow this layered reasoning:
Layer 1 — Parameter Validation & Sanity Checks
Before starting any negotiation, verify:
- Budget feasibility: Is
buyer_budget > 0andseller_base_price > 0? - ZOPA existence: Can the buyer's max concession ceiling ever reach the seller's min floor?
- Buyer ceiling at round N:
budget × (1 + 0.15 × (N-1)) - Seller floor at round N:
base_price × (1 - 0.12 × (N-1)) - If ceiling never ≥ floor within max_rounds → negotiation will always fail
- Buyer ceiling at round N:
- Intent clarity: Are
buyer_intentandseller_capabilityspecific enough for LLM reasoning?
Layer 2 — Strategy & Workflow Design
Choose the negotiation topology:
| Scenario | Recommended Approach |
|---|---|
| Simple price-only | Direct concession model (default) |
| Multi-dimensional (price + SLA + delivery) | Weighted utility scoring |
| Adversarial / untrusted parties | Add reputation threshold gate before negotiation |
| High-value transactions | Add human-in-the-loop approval before settlement |
Design the StateGraph workflow:
buyer_propose → seller_evaluate → [settle | buyer_decide | abort]
buyer_decide → [settle | seller_evaluate | abort]
Key routing decisions:
- Accept: Counterparty's offer meets or exceeds threshold → route to
settle - Counter: Gap exists but within concession range → route back to negotiation loop
- Reject/Timeout: Gap too large or max rounds exceeded → route to
abort
Layer 3 — Domain Constraints & Game Theory
Apply these domain-specific rules:
- Concession asymmetry: Buyer concedes faster (15%) than seller (12%) — this models real-world power dynamics where buyers typically have more alternatives
- ZOPA convergence: The Zone of Possible Agreement naturally emerges around round 3:
Round 1: Buyer $50.00 ←── gap ──→ Seller $80.00 Round 2: Buyer $57.50 ←─ gap ─→ Seller $70.40 Round 3: Buyer $65.00 ← → Seller $60.80 ← ZOPA overlap! - Reputation gating: PREFER filtering counterparties with
score < 0.3before entering negotiation - Settlement atomicity: x402 payment must be atomic — either fully settled or fully rolled back
Anti-Patterns
| Anti-Pattern | Problem | Correct Approach |
|---|---|---|
| Fixed-price offers | No convergence, wastes rounds | Use progressive concession with configurable rates |
| Unlimited rounds | Resource waste, no termination guarantee | Set max_rounds (default: 5) |
| Trusting all counterparties | Vulnerable to reputation manipulation | Gate on reputation.score ≥ threshold |
| Hardcoded concession rates | Inflexible across different deal sizes | Scale rates by deal magnitude or use adaptive rates |
| Skipping settlement verification | Payment may fail silently | Always check SettlementResult.status == "confirmed" |
| Mutable state in negotiation | Race conditions, debugging difficulty | Use frozen dataclasses, return new state per round |
Required Output Format
When reporting negotiation results, enforce this structure:
### Negotiation Trace
+-- Layer 1: [Parameter validation result]
| ↓
+-- Layer 2: [Strategy selected, workflow topology]
| ↓
+-- Layer 3: [Domain constraints applied, ZOPA analysis]
### Round-by-Round Log
| Round | Role | Action | Price | SLA | Reasoning |
|-------|--------|---------|---------|----------|-----------|
| 1 | buyer | propose | $50.00 | standard | ... |
| 1 | seller | counter | $75.00 | premium | ... |
| ... | ... | ... | ... | ... | ... |
### Settlement
- Status: [agreed/rejected/timeout]
- Final Price: $X.XX
- TX Hash: 0x...
- Reputation Update: buyer=X.XX, seller=X.XX
Environment Variables
| Variable | Required | Default | Description |
|---|---|---|---|
DASHSCOPE_API_KEY |
Yes | — | Qwen API key (get one) |
QWEN_MODEL |
No | qwen-plus |
Qwen model name |
Completeness Checklist
Before considering a negotiation implementation complete, verify:
- ZOPA feasibility checked before starting negotiation
- Max rounds configured and enforced
- Concession rates appropriate for deal size
- LLM JSON parsing has error handling with retry
- Settlement result status verified (not assumed)
- Reputation updated for both parties after settlement
- Failed negotiations also update reputation (with
success=False) - No hardcoded API keys in source code