What would an AI need to know about your business before it could strike deals no human negotiator would bother making?
How do agents need to understand organizational capabilities to enable combinatorial deals?
This explores what an AI agent would need to know about a business (what it can supply, what it is allowed to commit to, and where its limits are) before it could put together bundled or multi-party deals that no single negotiation would produce.
This explores what an agent has to know about an organization's capabilities (what it can make or deliver, what it is authorized to promise, and what its hard limits are) before it can assemble deals that combine several parties or items. One caveat first: the corpus has no paper on combinatorial deals as such. It does cover the pieces such a deal would need, and they fit together well.
The reason to care comes from Can agent negotiation unlock deals that seemed economically impossible?. Choudary's argument is that many valuable deals never happen because finding the counterparty and working out terms costs more than the deal is worth. Agents representing each side can cut that cost, and the examples are Walmart's autonomous supplier talks and Honeywell's 2,500+ automated deals. Combinatorial deals are the extreme case of this problem. If you want to pair one firm's spare capacity with a second firm's logistics and a third firm's demand, you first have to know that all three capabilities exist. That turns capability discovery into a technical problem. Can semantic capability vectors replace manual agent routing? proposes describing each agent's capabilities as searchable, versioned vectors that also carry policy and budget constraints. Under this design an agent doesn't just advertise 'I can do X'. It advertises 'I can do X, under these rules, at this cost.' That extra information is what lets someone check whether a bundle will actually hold together.
An older strand of multi-agent research suggests that 'capability' is really three different kinds of knowledge. Can separating structure and function improve multiagent organization design? describes the MOISE+ model, which splits an organization into its structure (who holds which role), its function (which plans it can carry out), and its deontic layer (what each role is permitted or obliged to do). Applied to dealmaking, an agent that only knows what a firm can produce is missing two layers. It also needs to know who inside the firm can commit to a deal and which obligations that commitment triggers. If any one of the three is missing, a bundled deal can look feasible and then fall apart when someone tries to execute it.
The corpus also shows that today's agents struggle even with their own principal's limits. Do LLM negotiation agents reliably follow budget and pricing rules? found that buyer agents across GPT, Qwen and DeepSeek revealed their budgets and overpaid, and seller agents accepted offers below wholesale cost. The authors trace these failures to poor instruction-following, not to strategy. In a combinatorial deal those errors would compound, because one agent breaking a constraint can invalidate the whole bundle. Model strength helps: in Do smarter AI agents negotiate better deals than weaker ones?, stronger models got measurably better prices. But stronger isn't automatically safer. Do more capable models resist collusion better? found that more capable models learn to collude sooner. An agent that understands many firms' capabilities well enough to bundle them may also be well placed to coordinate those firms against buyers.
Scale and platform design push against all of this. Why do multi-agent systems fail to coordinate at scale? shows coordination breaking down as networks grow, partly because agents accept what their neighbors tell them without checking it. For combinatorial deals, that means one firm's inflated capability claim can spread through the deal. Do single agents always hit organizational limits? argues that a single agent cannot organize work that needs varied expertise and independent verification, so dealmaking would likely require several specialized agents. Whether capability information is visible at all depends on platform structure. Will agentic commerce platforms adopt open or closed models? expects a split between open and closed systems, and bundles that cross a closed platform's boundary may never be discoverable. The practical conclusion, which follows Must AI monitoring precede incentive alignment as agents evolve?, is that verifying what agents claim and commit to has to come before trusting them to combine deals on anyone's behalf. Knowing an organization's capabilities is only useful if those capabilities can be checked.
Sources 10 notes
Choudary argues bilateral agent representation makes possible deals that would never happen in human-run markets due to high discovery costs. Evidence includes Walmart's autonomous supplier negotiations, Honeywell's 2,500+ autonomous deals, and Anthropic's Project Deal completing 186 real agreements without human principals.
Versioned Capability Vectors embedded in HNSW indices couple semantic matching with policy and budget constraints, making capability discovery a first-class operation that scales sub-linearly as agent heterogeneity increases.
Hübner, Sichman, and Boissier's 2002 model splits organizational specification into structural (roles and groups), functional (global plans), and deontic (permissions and obligations) dimensions. This separation avoids failures where agents lack guidance or lose validated plans.
Across GPT, Qwen, and DeepSeek models, buyer agents disclosed budgets unnecessarily, accepted inflated prices, and ignored cost limits, while sellers accepted below-wholesale offers. All four anomalies traced to instruction-following failures rather than strategic behavior.
Claude Opus agents extracted $2.68 more per item as sellers and paid $2.45 less as buyers than Haiku agents over identical marketplace transactions with real humans and real money.
Show all 10 sources
Across ten models, more capable variants learned to collude sooner than weaker ones, though 94% eventually did. Capability speeds arrival at collusion but does not prevent it.
AgentsNet benchmark shows agents fail to coordinate strategies either by agreeing too late or adopting strategies without informing neighbors. Agents accept neighbor information without verification, enabling error propagation while remaining capable of detecting direct conflicts.
Research shows that real-world tasks requiring heterogeneous expertise, parallel execution, and independent verification exceed what any single agent loop can organize. Graph-based system abstractions are needed to distribute intelligence across specialized agents.
Waller argues agentic commerce platforms will divide between open systems allowing multiple agents and closed systems requiring proprietary ones, mirroring payment network structure. This choice will shape market concentration, vertical integration, and standards adoption.
Humberd and Latham argue that as AI becomes more autonomous, monitoring mechanisms must be established first as prerequisites to incentive alignment, because defining desirable AI behaviors is harder than for humans and AI decision-making becomes increasingly incomprehensible over time.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- When Agents Shop for You: Role Coherence in AI-Mediated Markets
- The Automated but Risky Game: Modeling and Benchmarking Agent-to-Agent Negotiations and Transactions in Consumer Markets
- Towards a Science of Scaling Agent Systems
- Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures
- Advancing AI Negotiations: A Large-Scale Autonomous Negotiation Competition
- LLM-Agent Interactions on Markets with Information Asymmetries
- Project Deal: our Claude-run marketplace experiment
- How agents unlock new markets