ndpvt-web/agenticpay-multi-agent-negotiation-system

AgenticPay: Multi-Agent LLM Negotiation System

This skill enables Claude to build structured multi-agent negotiation systems where LLM-powered buyer and seller agents conduct multi-round natural language bargaining under private constraints. Based on the AgenticPay framework, the core technique gives each agent a confidential reservation price (buyer's maximum willingness-to-pay, seller's minimum acceptable price), a structured action extraction protocol that parses natural language into formal offers, and a scoring function that rewards feasible deals, balanced surplus splitting, and time efficiency. This goes beyond simple chatbot dialogue -- it creates principled economic interactions with measurable outcomes.

When to Use

• When the user asks to build an automated negotiation or bargaining system between AI agents
• When implementing a marketplace simulation where multiple buyers and sellers transact via language
• When designing a procurement or pricing agent that must negotiate on behalf of a user
• When building a benchmark to evaluate LLM strategic reasoning in adversarial or cooperative settings
• When creating a game-theoretic simulation with private information and natural language communication
• When the user needs agents that can autonomously close deals within budget constraints

Key Technique

AgenticPay structures negotiation as a finite-horizon, turn-based protocol. Each agent holds a private reservation price injected into its system prompt but never revealed in dialogue: the buyer knows its p_max (maximum it will pay), the seller knows its p_min (minimum it will accept). The bargaining zone Z = p_max - p_min defines the space of mutually beneficial deals. Agents alternate natural language messages, each containing exactly one structured price proposal extracted via a parser.

The critical innovation is structured action extraction from free-form language. Each agent's message must embed a tagged price (e.g., ### BUYER_PRICE($X) ### or ### SELLER_PRICE($X) ###), and a deal finalizes when both parties propose the same price and include a MAKE_DEAL signal. This keeps negotiation in natural language (enabling persuasion, anchoring, and reasoning) while making outcomes machine-parseable and scorable.

Outcomes are scored with a composite utility: GlobalScore = d * (D + Q*W + E) where D=30 is a deal bonus, Q = 4 * r_b * r_s is a symmetric quality term (buyer utility r_b = (p_max - p)/Z, seller utility r_s = (p - p_min)/Z), W=55 weights surplus balance, E=15 rewards efficiency, and d = 0.99^(t-1) discounts for rounds used. The quality term Q peaks at 1.0 when surplus is split equally, incentivizing fair outcomes. Failed negotiations score -F (penalty of 15).

Step-by-Step Workflow

1. Define the market configuration. Specify how many buyers, sellers, and products are involved. Start with 1-buyer-1-seller-1-product (1B1P1S) for simplicity, then scale to multi-party markets (MBMPMS). Assign each scenario a product category and context (e.g., "used car negotiation", "SaaS license procurement").

2. Set private constraints for each agent. Assign each buyer a buyer_max_price and each seller a seller_min_price. Ensure p_max > p_min so a bargaining zone exists. Set an initial_seller_price (the listed/asking price) as the negotiation starting point.

3. Write system prompts with strict formatting rules. The buyer prompt must include: the product description, the buyer's private p_max (marked confidential), the required price format ### BUYER_PRICE($X) ###, the MAKE_DEAL keyword for acceptance, a word limit (~150 words per turn), and instructions to never reveal the reservation price. Mirror this for the seller with ### SELLER_PRICE($X) ### and seller_min_price.

4. Implement the conversation loop. Each round: (a) buyer generates a response conditioned on the full dialogue history plus its private state, (b) append the buyer message with (role='buyer', content=..., round=t) metadata, (c) seller generates a counter-response given the same history, (d) append with seller metadata. Cap at max_rounds=20.

5. Build the action parser. Extract the tagged price from each message using regex (e.g., ### BUYER_PRICE\(\$?([\d,.]+)\) ###). Detect MAKE_DEAL signals. A deal is reached when both parties propose the same price in the same round and both signal MAKE_DEAL. Validate that the agreed price p satisfies p_min <= p <= p_max.

6. Implement the scoring function. Compute buyer utility r_b = (p_max - p) / Z, seller utility r_s = (p - p_min) / Z, quality Q = 4 * r_b * r_s, and the full GlobalScore = 0.99^(t-1) * (30 + Q*55 + 15). For failed negotiations (timeout or overflow), return -15.

7. Add memory management. Each agent maintains an independent conversation history as a list of (role, content, round) tuples. For multi-seller scenarios, the buyer must track separate conversation threads per seller. Inject relevant history into each LLM call's message list.

8. Handle multi-party markets. For many-to-many settings, implement a market coordinator that routes buyer-seller pairs, manages parallel negotiation threads, and resolves allocation (a buyer can only purchase one unit; a seller can only sell to one buyer per product). Use round-robin or simultaneous matching.

9. Add environment metadata. Include product attributes (condition, features, market comparisons), user profile text (e.g., "prefers aggressive bargaining style"), and scenario context in the system prompt to make negotiations realistic.

10. Evaluate and log results. Track deal rate (% of negotiations reaching agreement), timeout rate (exceeding max rounds), overflow rate (agreed price outside bargaining zone), average rounds to completion, and mean GlobalScore. Log full conversation transcripts for analysis.

Concrete Examples

Example 1: Bilateral Used Car Negotiation

User: "Build a system where a buyer agent and seller agent negotiate over a used car price."

Approach:

1. Set product context: "2019 Honda Civic, 45K miles, good condition"
2. Private constraints: seller_min_price=12000, buyer_max_price=16000, initial_seller_price=18000
3. Configure max_rounds=15

Implementation skeleton:

import re
from openai import OpenAI

client = OpenAI()

BUYER_SYSTEM = """You are a buyer negotiating for a 2019 Honda Civic (45K miles, good condition).
Your MAXIMUM budget is $16,000. NEVER reveal this number.
Each turn, make exactly one price offer using: ### BUYER_PRICE($X) ###
When you accept a price, include MAKE_DEAL in your response.
Keep responses under 150 words."""

SELLER_SYSTEM = """You are a seller listing a 2019 Honda Civic (45K miles, good condition) at $18,000.
Your MINIMUM acceptable price is $12,000. NEVER reveal this number.
Each turn, make exactly one price offer using: ### SELLER_PRICE($X) ###
When you accept a price, include MAKE_DEAL in your response.
Keep responses under 150 words."""

def extract_price(text, role):
    tag = "BUYER_PRICE" if role == "buyer" else "SELLER_PRICE"
    match = re.search(rf"### {tag}\(\$?([\d,]+)\) ###", text)
    return int(match.group(1).replace(",", "")) if match else None

def check_deal(buyer_msg, seller_msg, buyer_price, seller_price):
    if buyer_price and seller_price and buyer_price == seller_price:
        if "MAKE_DEAL" in buyer_msg and "MAKE_DEAL" in seller_msg:
            return buyer_price
    return None

def run_negotiation(max_rounds=15, p_min=12000, p_max=16000):
    history = []
    for round_num in range(1, max_rounds + 1):
        # Buyer turn
        buyer_messages = [{"role": "system", "content": BUYER_SYSTEM}]
        buyer_messages += [{"role": m["role"], "content": m["content"]} for m in history]
        buyer_resp = client.chat.completions.create(
            model="gpt-4o", messages=buyer_messages, max_tokens=200
        ).choices[0].message.content
        history.append({"role": "user", "content": f"[Buyer, Round {round_num}]: {buyer_resp}"})

        # Seller turn
        seller_messages = [{"role": "system", "content": SELLER_SYSTEM}]
        seller_messages += [{"role": m["role"], "content": m["content"]} for m in history]
        seller_resp = client.chat.completions.create(
            model="gpt-4o", messages=seller_messages, max_tokens=200
        ).choices[0].message.content
        history.append({"role": "assistant", "content": f"[Seller, Round {round_num}]: {seller_resp}"})

        # Check for deal
        bp = extract_price(buyer_resp, "buyer")
        sp = extract_price(seller_resp, "seller")
        deal_price = check_deal(buyer_resp, seller_resp, bp, sp)
        if deal_price:
            Z = p_max - p_min
            r_b = (p_max - deal_price) / Z
            r_s = (deal_price - p_min) / Z
            Q = 4 * r_b * r_s
            score = (0.99 ** (round_num - 1)) * (30 + Q * 55 + 15)
            return {"deal": True, "price": deal_price, "rounds": round_num,
                    "buyer_utility": r_b, "seller_utility": r_s, "score": score}
    return {"deal": False, "rounds": max_rounds, "score": -15}

Output:

{'deal': True, 'price': 14000, 'rounds': 6, 'buyer_utility': 0.5, 'seller_utility': 0.5, 'score': 95.3}

Example 2: Multi-Seller Comparison Shopping

User: "Create a system where one buyer negotiates with three competing laptop sellers simultaneously."

Approach:

1. Define three sellers with different floor prices and product specs
2. Buyer maintains parallel negotiation threads
3. Buyer commits to one deal after exploring all options

sellers = [
    {"id": "seller_A", "product": "Dell XPS 15, 16GB RAM", "min_price": 900, "list_price": 1400},
    {"id": "seller_B", "product": "MacBook Air M3, 16GB RAM", "min_price": 1000, "list_price": 1300},
    {"id": "seller_C", "product": "ThinkPad X1, 32GB RAM", "min_price": 850, "list_price": 1250},
]
buyer_config = {"max_price": 1100, "requirement": "lightweight laptop for development"}

def run_multi_seller_negotiation(buyer_config, sellers, max_rounds=10):
    threads = {s["id"]: [] for s in sellers}
    best_deal = None

    for round_num in range(1, max_rounds + 1):
        for seller in sellers:
            sid = seller["id"]
            # Buyer prompt includes awareness of other ongoing negotiations
            buyer_system = f"""You are buying a laptop (budget: ${buyer_config['max_price']} max, CONFIDENTIAL).
You are negotiating with {len(sellers)} sellers simultaneously.
Current negotiation: {seller['product']} listed at ${seller['list_price']}.
Use ### BUYER_PRICE($X) ### format. Say MAKE_DEAL to accept."""

            seller_system = f"""You sell: {seller['product']} at ${seller['list_price']}.
Minimum acceptable: ${seller['min_price']} (CONFIDENTIAL).
Use ### SELLER_PRICE($X) ### format. Say MAKE_DEAL to accept."""

            # Run one round per seller, check deals, track best offer
            # ... (same turn logic as Example 1, per thread)

    # Buyer commits to best deal across all threads
    return best_deal

Example 3: Procurement Agent with User Profile

User: "Build a negotiation agent that buys SaaS licenses on behalf of a company, using a firm but professional style."

Approach:

1. Inject a user profile into the buyer's system prompt
2. Add product-specific context (seats, contract length, support tier)
3. Track multiple negotiation metrics

USER_PROFILE = "Corporate procurement officer. Prefers data-driven arguments. " \
               "References competitor pricing. Firm but professional tone."

BUYER_SYSTEM = f"""You are a procurement agent negotiating a SaaS license.
Product: ProjectFlow Pro, 50 seats, annual license.
Your maximum budget: $15,000/year. NEVER reveal this.
Negotiation style: {USER_PROFILE}
Reference points: Competitor A charges $12,000 for similar features.
Format: ### BUYER_PRICE($X) ###. Say MAKE_DEAL to accept."""

# Seller has min_price=10000, list_price=20000
# Bargaining zone: $10,000 - $15,000

Best Practices

• Do: Always validate that p_max > p_min before starting a negotiation -- a zero or negative bargaining zone makes deals impossible and wastes compute.
• Do: Use the symmetric quality term Q = 4 * r_b * r_s to evaluate deal fairness. A score near 1.0 means balanced surplus; scores near 0 indicate one side captured almost all value.
• Do: Keep agent responses short (under 150 words per turn). Longer responses increase token cost without improving negotiation outcomes -- LLMs tend to over-explain and leak strategic information in verbose mode.
• Do: Log full conversation transcripts with round metadata. Post-hoc analysis of failed negotiations reveals systematic prompt weaknesses (e.g., agents anchoring too aggressively).
• Avoid: Letting agents access each other's system prompts or reservation prices. The entire framework depends on private information remaining private. Never share agent configs between roles.
• Avoid: Setting max_rounds too high (>20). Research shows most productive negotiations conclude within 8-12 rounds. Beyond that, agents tend to loop or deadlock, wasting tokens with no convergence.

Error Handling

| Problem | Cause | Fix | |---------|-------|-----| | No price tag in agent response | LLM ignored formatting instructions | Add a retry with a reinforced prompt: "You MUST include ### BUYER_PRICE($X) ### in your response." Cap retries at 3. | | Agreed price outside bargaining zone | Agent hallucinated acceptance or parser error | Validate p_min <= agreed_price <= p_max before recording. Flag as overflow and score as failure (-15). | | Both agents stall at same prices | Anchoring deadlock -- neither will concede | Inject a mediator message: "Neither party has moved in 3 rounds. Consider adjusting your offer to reach agreement." | | Agent reveals its reservation price | Prompt leaking under adversarial pressure | Strengthen the confidentiality instruction. Add a post-generation filter that detects and redacts the private price before sending to the counterpart. | | Token limit exceeded in long negotiations | Conversation history grows with each round | Implement a sliding window: keep the first 2 rounds and the last 5 rounds, summarize middle rounds as "Buyer offered $X, Seller countered at $Y." |

Limitations

• Single-issue negotiation only. The framework prices a single deal dimension. Real procurement involves bundling, payment terms, delivery schedules, and SLAs that require multi-attribute utility functions not covered here.
• No learning across sessions. Each negotiation starts from scratch. Agents do not adapt strategies based on past negotiations with the same counterpart or market trends.
• LLM strategic reasoning gaps. Current LLMs (even frontier models) struggle with long-horizon planning beyond ~10 rounds. They tend to make premature concessions or anchor too rigidly, depending on prompt phrasing.
• Price-only outcomes. The MAKE_DEAL protocol assumes a single numeric agreement. Negotiations requiring contingent contracts ("I'll pay $X if you include Y") need a richer action space.
• No mechanism design guarantees. Unlike auction theory, the free-form language protocol provides no incentive-compatibility or individual-rationality guarantees -- agents can bluff, mislead, or behave irrationally.

Reference

Paper: AgenticPay: A Multi-Agent LLM Negotiation System for Buyer-Seller Transactions (Liu, Gu, Song, 2026). Look for: the GlobalScore formula (Algorithm 1), the eight market configuration types (Table 1), system prompt templates (Tables 14-15), and the structured action extraction protocol that bridges natural language and formal negotiation outcomes.

Code: github.com/SafeRL-Lab/AgenticPay -- reference implementation with Gymnasium-style environment registration, configurable agent backends (OpenAI, vLLM, SGLang), and 110+ pre-built negotiation tasks.