Why the Agent Economy may not end in autonomous A2A.
One common vision of the Agent Economy is that every company has an AI agent. Those agents find each other through protocols like A2A, negotiate, buy, sell, and close deals on their own.
But allowing agents to communicate across companies is not the same as authorizing them to make commercial commitments.
By fully autonomous A2A, I mean agents on both sides making those commitments on their own, within limits their companies set, without a person approving each deal.
I don't think this will become the default model for the Agent Economy, for three reasons:
- Cross-company coordination doesn't require an agent at every company. A buyer, a provider, or a marketplace can run the coordinating agent, while other companies offer APIs or keep their agents internal.
- The two sides don't necessarily have the same reason to automate. One company may gain enough to run a negotiating agent while the other has little reason to offer one to its counterpart.
- Needing independent judgment doesn't mean companies can delegate it to AI. Strategic negotiations make a particularly strong case for independent decision-making on both sides. Yet those decisions depend on context that is hard to capture and priorities the companies may not have settled internally.
Two Agents Are Not the Default
1. The Client Uses an Agent, the Provider Uses an API
Say I have an AI travel assistant that knows my budget, schedule, and how much more I will pay for a direct flight. It decides what counts as a good flight for me.
For a standard booking, the airline can return flight options, fares, and seat availability through an API, and accept bookings the same way.
The airline may use agents heavily inside its own systems and still offer a stable API to external partners.
2. The Customer Uses the Provider’s Agent
A bank, insurer, healthcare service, or B2B provider may need to ask follow-up questions, check private records, apply internal rules, and handle unusual cases.
Customers can use the provider's agent directly. They don't need an agent of their own.
3. The Marketplace Uses the Agent
In freight, the agent may belong to the marketplace instead of to every company.
One shipment might involve a truck to the port, an ocean carrier, and another truck at the destination. A buyer may want to compare many carriers, split the shipment across providers, or let a smaller carrier compete for just one part of the journey.
A marketplace can collect the buyer’s needs, compare offers, and put together a shipment plan.
Carriers can provide quotes, capacity information, and booking services through APIs without exposing agents of their own.
The marketplace can request quotes without needing access to each carrier's internal pricing rules or commercial limits.
A2A can lower coordination costs. But if a carrier still earns more by selling the whole journey, lower costs alone won't make it sell individual legs.
In fragmented markets like freight, I expect marketplaces to have a stronger economic case for running the coordinating agent than individual carriers do, since they can spread the cost across many buyers and carriers.
The Economics Often Favor One-Sided Automation
Negotiation looks like a natural A2A use case because both sides have their own interests and decisions to make.
But that does not mean both sides have the same economic reason to automate.
In the 2022 Harvard Business Review case, Walmart had more than 100,000 suppliers. Many smaller supplier relationships weren't worth assigning a human buyer to negotiate individually, but small savings across thousands of suppliers could add up. That scale gave Walmart a strong reason to automate.
Walmart used AI-powered negotiation software with procurement goals and acceptable trade-offs set in advance. Suppliers negotiated through a chatbot.
Suppliers still made their own decisions. They could respond to Walmart's chatbot without deploying negotiating agents of their own.
Scale can make one-sided automation the better design: both sides need to make decisions, but only one side may have a strong reason to automate them.
The Narrow Sweet Spot for Fully Autonomous A2A
For fully autonomous A2A to make sense, a negotiation must be complex enough for two AI agents to add real value, yet clear enough that both companies are willing to let those agents close deals without human approval each time.
Routine negotiations are easier to automate. Companies can define pricing limits, acceptable trade-offs, deadlines, and approval rules in advance. Once those boundaries are clear, conventional software can handle much of the negotiation. LLMs can still help with language and exceptions, but two agents reasoning freely may add little value.
Consider a major supply agreement involving pricing, capacity, delivery commitments, and long-term business risks. Both companies must evaluate offers against their own priorities, which may differ and remain private.
If both companies handle enough complex negotiations to justify their own agents, A2A becomes a plausible choice. Each agent could evaluate offers, negotiate terms, and make commitments on its company's behalf.
But this is also where full autonomy becomes hardest.
Why Strategic Judgment Is Hard to Define in Advance
Strategic decisions depend on more than information stored in company systems. Some judgments are difficult even for the people making them to explain:
- Experience and intuition: A senior salesperson may sense that a customer is losing confidence without knowing exactly why.
- Relationships and trust: A company may accept less favorable terms because a supplier supported it during a difficult period.
An LLM can analyze records, ask questions, and infer preferences. But those inferences may not reflect what someone is actually willing to accept.
There is a deeper problem: sometimes the company hasn't even decided what it wants.
Suppose Sales wants to offer a discount to keep a major customer while Finance wants to protect profit. Both departments may have clear objectives, but the company has not decided which should take priority.
That decision may emerge through discussion. Leadership might reconsider the customer's long-term value, weigh the risk of losing the account, or change its priorities after hearing both sides. The discussion does not simply reveal an existing company position. It helps create one.
The same problem arises when circumstances change. A company may normally prioritize profit but reconsider when a key supplier fails or an unexpected market opportunity appears. Not every future situation can be anticipated, and companies may deliberately leave some trade-offs open until they understand the circumstances.
An agent cannot find a decision the company itself has not made.
At that point, the agent isn't retrieving an answer. It must either help the company reach a decision or be authorized to make one.
Decision-Making Is Not Decision Authority
In major deals, Procurement, Finance, Legal, and senior management may deliberately hold separate decision rights to manage risk and accountability. An agent may identify an attractive deal without having permission to commit the company.
A company could give its agent clear goals, limits, and authority. But setting those boundaries is itself a business decision. The company must decide which trade-offs to settle in advance and which decisions it is willing to delegate.
That leaves two approaches:
- Set tighter boundaries in advance. The agent can negotiate and commit within those limits. But the more important trade-offs the company settles beforehand, the less it needs open-ended AI judgment during the negotiation, and the more conventional software can handle.
- Give the agent broader discretion. It can handle unexpected situations and decide trade-offs the company has not fully specified. But that means allowing it to make business decisions the company itself may not have settled. If the company is unwilling to delegate those decisions, people must remain involved before major commitments are made.
This is what makes the middle ground so narrow. Every important trade-off settled in advance reduces what the agent needs to decide. Every consequential trade-off left open requires the company to delegate that decision or retain human approval.
And fully autonomous A2A requires both companies to accept that arrangement.
Better models can improve the reasoning. They cannot retrieve a position the company hasn't formed or grant themselves authority the company hasn't given.
Fully autonomous A2A is squeezed between the two.
Fully Autonomous A2A Is Not a Maturity Level
Deloitte’s B2B agentic-commerce model puts fully autonomous cross-company A2A at the top of its maturity ladder, although it notes that different sectors may stop at different stages.
I think that mixes two different questions. The earlier stages measure how much work a company automates. The final stage is about how two companies choose to interact.
A company can automate deeply inside its own operations and still use APIs, marketplaces, provider-owned agents, or people working with agents when dealing with other companies.
That does not make it less mature.
Fully autonomous A2A may fit some markets, but it shouldn't be the assumed destination.
Before putting agents on both sides of a deal, I would ask four questions:
- Value: Who gains enough to justify running an agent?
- Context: Who holds the information the agent needs, and what can they share?
- Judgment: Which decisions require more than fixed rules, and who has the authority to make them?
- Coordination: Could an API, the provider, a marketplace, or one side handle it instead?
The answers may lead to two agents, one agent, or no externally exposed agents at all.
Markets do not exist to complete a technology roadmap.
