Anthropic reportedly works with lawyers to establish negotiation rules for autonomous AI agents

Anthropic is reportedly consulting legal experts and law firms to determine how autonomous AI agents should negotiate on behalf of human users, as artificial intelligence systems move toward independently handling purchases, commercial disputes, contract discussions and potentially salary negotiations.

The development, reported on October 18, 2026, concerns the rules governing interactions between agents with competing objectives. An AI representing a buyer may seek the lowest possible price, while another representing a seller attempts to maximize revenue. Both may be acting legitimately, yet their interactions raise questions about authorization, bargaining tactics, confidentiality, fairness and contractual responsibility.

Anthropic is examining how human legal professionals represent clients, including how much discretion an agent should receive, when it should seek additional instructions and what constraints should apply when dealing with another autonomous system. The consultations have not been accompanied by a public product launch, finalized negotiation standard or commercial deployment schedule.

The initiative builds on Anthropic's research into AI-mediated markets, particularly Project Deal and Project Swap, which examined how Claude-powered agents negotiate, exchange goods and represent human preferences. In Project Swap, involving 201 Anthropic employees, agents could negotiate effectively under controlled conditions but frequently lacked adequate information about users' preferences.

··········

ANTHROPIC IS EXPLORING HOW AI AGENTS SHOULD NEGOTIATE ON BEHALF OF PEOPLE.

The reported legal consultations address a problem conventional conversational systems rarely encounter: autonomous agents can pursue objectives directly conflicting with those of other agents. A shopping assistant might seek the lowest price while a retailer's agent protects margins. Both goals may be legitimate, but negotiations require rules about permissible tactics and binding commitments.

........

Area	Question for autonomous negotiation
User authorization	What may the agent negotiate or accept?
Negotiation strategy	How aggressively may it pursue better terms?
Disclosure	What information may it reveal?
Conflicts of interest	Whose interests must it prioritize?
Human oversight	When must it request approval?
Contract formation	When does an accepted offer become binding?
Identity	How can counterparties verify authority?
Accountability	Who is responsible for an unauthorized agreement?
Dispute resolution	How can a transaction be challenged or reversed?

........

An agent must understand preferences, evaluate trade-offs, maintain constraints and distinguish a provisional offer from a final commitment. A technically successful negotiation can still be undesirable if it optimizes an objective the user never intended to prioritize.

··········

LEGAL EXPERTISE COULD HELP DEFINE THE LIMITS OF AGENT AUTHORITY.

Human representatives typically operate under instructions defining acceptable prices, confidential information, approval conditions and withdrawal circumstances. Autonomous agents need comparable boundaries, but translating legal and commercial instructions into machine-executable behavior is difficult.

Consider an agent negotiating a software subscription. Its user may authorize a price reduction but prohibit automatic renewal, personal-data transfers or commitments longer than twelve months. A model optimizing primarily for price could obtain a substantial discount while overlooking unfavorable contractual provisions.

Effective authorization must represent multiple constraints rather than a single instruction to secure the best deal. One approach reportedly under consideration involves user-signed documents defining the agent's mandate. Such documents could establish evidence of authorization, but legal effectiveness depends on jurisdiction, transaction type and implementation. A signed mandate does not automatically make every AI action enforceable.

··········

PROJECT SWAP PROVIDES DIRECT EVIDENCE ABOUT AI-TO-AI NEGOTIATION.

In research published September 24, Project Swap involved 201 employees across six Anthropic offices in a controlled book-exchange market. Participants supplied a book and discussed reading preferences with a Claude-powered agent, which then negotiated with other agents in a digital marketplace.

........

Project Swap metric	Reported result
Participants	201
Participating offices	6
Pairwise preference agreement	61%
Random preference baseline	50%
Theoretical maximum allocation efficiency	0.89
Decentralized market efficiency	0.55
Best allocation using inferred preferences	0.60
Average willingness to delegate annual book budget	Approximately 30%

........

The 61% agreement measures how often Claude correctly predicted which of two books a participant preferred; it does not mean that 61% of transactions succeeded. Researchers evaluated both preference representation and negotiation outcomes. Agents generally negotiated competently with the information available, but that information was often incomplete.

··········

UNDERSTANDING USER PREFERENCES WAS A GREATER LIMITATION THAN NEGOTIATION SKILL.

Anthropic decomposed the gap between theoretically optimal allocation and actual decentralized negotiation. Using participants' actual rankings, the theoretical optimum scored 0.89. An allocation using Claude-inferred preferences achieved 0.60 when evaluated against real preferences, while the decentralized agent market reached 0.55.

........

Allocation scenario	Efficiency score
Theoretical optimum using actual preferences	0.89
Best allocation using Claude-inferred preferences	0.60
Decentralized agent negotiations	0.55

........

The total shortfall was 0.34. Of this, 0.29 arose before decentralized bargaining because agents inadequately represented preferences; the remaining 0.05 reflected decentralized market execution. Anthropic attributed approximately 85% of the shortfall to preference representation and 15% to market execution.

Those proportions describe this experiment, not a universal distribution of errors. The implication is that improving bargaining strategy may deliver limited value if an agent misunderstands its principal. For consequential transactions, structured preference gathering may matter more than additional negotiating power.

··········

STRONGER MODELS IMPROVED MARKET OUTCOMES MORE THAN AGGRESSIVE INSTRUCTIONS.

Project Swap compared Claude models and negotiation instructions. When market efficiency was evaluated against inferred preferences, stronger models generally produced better allocations.

........

Agent model	Reported market efficiency
Claude Haiku 4.5	0.75
Claude Sonnet 4.5	0.80
Claude Opus 4.8	0.88
Claude Fable 5	0.86

........

These are experimental market outcomes, not universal negotiation benchmark scores. The results were not strictly monotonic: Fable 5 scored slightly below Opus 4.8.

Researchers also compared aggressive negotiation instructions with prompts encouraging consideration of other participants' welfare. More aggressive agents achieved a modest individual advantage of approximately 0.02 on the experiment's preference-based efficiency scale. That difference was relatively small compared with some model-upgrade effects and does not establish that aggressive bargaining is universally optimal.

··········

NEGOTIATION RULES MUST BALANCE USER LOYALTY WITH FAIR MARKET BEHAVIOR.

Agents may need to balance immediate user benefit against time, relationships and counterparties' legitimate interests. A user seeking a refund may prefer assertive bargaining, while a company negotiating a long-term supplier relationship may prioritize continuity.

........

Negotiation strategy	Potential advantage	Potential risk
Aggressive bargaining	Better immediate terms	Delays, disputes, relationship damage
Cooperative bargaining	Faster agreement, reciprocal benefits	Unnecessary concessions
Strict reservation limits	Prevents unacceptable deals	Misses mutually beneficial compromises
Adaptive bargaining	Responds to counterpart behavior	Difficult to audit or predict
Human approval checkpoints	Preserves user control	Adds latency and operational cost

........

The appropriate strategy depends on user preferences and consequences of failure. An agent should not assume that maximizing one financial variable captures the full economic value of an agreement.

··········

CONFIDENTIAL INFORMATION BECOMES A STRATEGIC VULNERABILITY.

Negotiating agents may know a user's maximum budget, salary, competing offers or willingness to compromise. Revealing those details can weaken bargaining position even when the information is accurate.

Project Swap found agents sometimes disclosed their preferences, though generally not complete rankings. Between 78% and 96% of agents, depending on the model, mentioned their top-ranked book during negotiations. Outright false statements about top preferences were uncommon.

Truthful disclosure can still be strategically harmful. Information-sharing policies may need to operate at the data-access and tool-permission level, not merely through natural-language instructions. This becomes particularly important in procurement, salary discussions and commercial contracting.

··········

AI-TO-AI NEGOTIATIONS INTRODUCE NEW SECURITY AND MANIPULATION RISKS.

Messages from counterparties are untrusted input. A malicious counterparty could insert instructions attempting to override a buyer agent's priorities or extract confidential information through prompt injection.

Defenses require separating external negotiation content from higher-priority user instructions and system constraints. Technical safeguards may include restricted tool permissions, structured offers, contract validation and explicit authorization checks before execution.

These controls become critical when agents can access payment systems, sign documents, update records or transfer funds. Unrestricted free-form messaging without separating instructions from data creates avoidable security exposure.

··········

CONTRACT FORMATION REQUIRES MORE THAN AN AGENT SAYING YES.

A negotiation becomes legally consequential when an agent accepts an agreement. Systems must distinguish inquiries, nonbinding proposals, conditional acceptances and final commitments. Conversational statements alone may be ambiguous without structured terms.

........

Stage	Agent responsibility	Recommended control
Mandate	Understand user objectives	Explicit scope and limits
Discovery	Identify counterparties	Identity verification
Negotiation	Exchange proposals	Confidentiality and policy checks
Evaluation	Compare offers	Constraint validation
Approval	Obtain required consent	Human or delegated authorization
Commitment	Accept final terms	Signed, structured transaction
Execution	Fulfill agreement	Restricted transactional tools
Audit	Preserve evidence	Tamper-resistant records

........

This is a proposed operational architecture, not an Anthropic product specification or announced legal standard. Legal consequences depend on contract law, consumer protection, agency principles and authorization. A digital signature alone cannot resolve jurisdictional differences.

··········

SALARY NEGOTIATIONS CREATE PARTICULARLY SENSITIVE REQUIREMENTS.

Salary bargaining requires evaluating base compensation, bonuses, equity, vesting, benefits, remote-work terms and termination provisions. A higher salary may be less valuable than a lower offer with stronger equity or contractual protections.

An agent must distinguish authorized disclosures from private circumstances. Employment negotiations also raise concerns about discrimination, information asymmetry and employers inferring candidates' minimum acceptable compensation.

Employer and candidate agents may optimize different objectives, potentially producing outcomes neither person fully understands. Anthropic's reported interest should not be interpreted as confirmation of a commercially available Claude salary-negotiation product.

··········

ECONOMIC BENEFITS DEPEND ON TRANSACTION VALUE AND NEGOTIATION COST.

Autonomous agents could make negotiation economical for transactions too small to justify human labor. Yet inference, tools, repeated bargaining rounds and human approval still create costs.

Consider a hypothetical $600 annual subscription. A 10% discount saves $60. With $2 in agent costs and $3 in human-review costs, the net saving is $55.

........

Illustrative transaction	Amount
Original annual subscription	$600
Negotiated discount	10%
Gross saving	$60
Agent inference and tools	$2
Human-review cost	$3
Net saving	$55

........

These are illustrative assumptions, not Anthropic experimental results. Economics worsen when negotiations fail, consume many model calls or create contractual risk. A full evaluation should measure net realized value, including disputes, compliance and unsuccessful transactions, rather than the nominal discount alone.

··········

UNEQUAL AGENT CAPABILITIES COULD CREATE NEW MARKET ADVANTAGES.

Project Swap found stronger models could outperform weaker models in mixed-agent markets. Organizations with sophisticated models, specialized tools and proprietary transaction histories may gain advantages over consumers using inexpensive assistants.

Those advantages might arise from better preference modeling, more computation or superior market information. Conversely, accessible AI agents could give ordinary consumers negotiation assistance previously reserved for paid professionals.

The eventual distributional effects depend on model pricing, interoperability, market rules and access. Controlled experimental results do not establish which participants will benefit most in real markets.

··········

MARKETPLACES WILL NEED IDENTITY, AUDITABILITY AND DISPUTE RESOLUTION.

Project Swap operated in a controlled environment where Anthropic knew the human associated with every agent. Commercial marketplaces are more complex: agents may operate through third-party providers, delegated credentials or unfamiliar systems.

Identity mechanisms could establish authority and accountability. Transaction logs could preserve proposals, approvals, final terms and execution records. Dispute procedures are necessary when an agent accepts unauthorized terms, a counterparty fails to deliver or a transaction is challenged.

Anthropic's research noted that even its controlled book exchange encountered fulfillment problems: participants did not always deliver books their agents had agreed to exchange. Reaching an agreement and completing a transaction are separate capabilities.

··········

THE REPORTED CONSULTATIONS DO NOT YET ESTABLISH A COMMERCIAL PRODUCT.

The October 18 reporting concerns consultations with legal professionals. It does not establish that Anthropic has finalized a universal negotiation protocol, published binding behavioral rules or launched a consumer service for autonomous salary and contract bargaining.

Earlier research demonstrates that Claude agents can conduct structured negotiations and that their behavior can be studied experimentally. Project Swap also reveals limitations in preference inference, information disclosure and fulfillment.

These findings support further research, not unrestricted deployment in financially consequential environments. Organizations should assess authorization, preference accuracy, security, transaction reliability and legal requirements before adopting such systems.

··········

ANTHROPIC'S NEXT CHALLENGE IS DEFINING WHAT AN AI AGENT IS AUTHORIZED TO DO.

Anthropic's reported engagement with legal experts reflects a transition from AI generating recommendations to AI acting for users in transactions with other autonomous systems.

The research suggests agents can negotiate effectively under controlled conditions, but their performance is constrained by imperfect understanding of human preferences and by market rules.

Project Swap's central quantitative result was that approximately 85% of the observed efficiency shortfall arose from imperfect preference representation, rather than decentralized negotiation itself. Better negotiation models must therefore be paired with user authorization, confidentiality controls, structured commitments and dispute resolution.

The defining challenge for autonomous AI negotiation is not simply teaching agents to secure better deals. It is ensuring that the deals reflect what users actually want, stay within authorized boundaries and remain accountable.

··········

FOLLOW US FOR MORE.

DATA STUDIOS

datastudios.org