Agent Driven Commerce

I've written before about how AI agents will become a meaningful component of all online commerce:

“Outsourcing purchasing power to agents is a natural evolution from how we discover and buy new products today. As AI agents become net new market participants, they free us from the mess of ad-driven faux optionality and let new products compete on equal footing.”

- AI Agents as a new distribution channel (Feb 2024)

Since then, many of the pieces to realizing this vision have fallen into place.

The Agent-Driven Commerce Stack

We can think of the agent commerce stack as having three layers:

• intent: understanding what users actually want (informed by personalization)
• discovery: searching the sample space of all possible inventory
• execution: making a purchase within appropriate guardrails

While the first two capabilities have existed in various forms, Stripe’s Agent Toolkit has now completed the picture by enabling payment processing in LLM workflows. It works simply: use Stripe Issuing to create a pre-filled virtual card and programmatically accept or reject payments.

Some companies, like Skyfire, achieve similar results using stablecoins. Most agent implementations operate with predefined budgets, which serve as crucial guardrails against potential overspending – especially important given common concerns about LLM agent reliability.

Now that we have the implementation details sorted. What kind of agent interaction patterns will we see in the wild? I can think of two big buckets:

1. Agent-to-human transactions: agents interact with real-world human systems to make purchases. This is a great fit for services industries that aren’t fully digital. Imagine an AI travel agent that not only books your flights and hotels but also coordinates with local human tour guides. This removes the friction of dealing with time zone differences, language barriers, and the uncertainty of finding reliable guides in unfamiliar places. The human guide still provides irreplaceable in-person service, but the AI agent handles all the coordination and transaction complexity.
2. Agent-to-agent transactions: in fully digital industries, agents will transact with other agents. Consider a B2B example: A procurement agent negotiates with multiple supplier agents to maintain optimal inventory levels. The agent monitors stock levels, predicts demand using historical data, and automatically negotiates purchases with supplier agents. What makes this powerful is the continuous optimization: the store's agent could adjust order quantities based on real-time sales data, while supplier agents could dynamically adjust prices based on their inventory and production capacity. This creates a more efficient market than traditional bulk ordering on fixed schedules. The entire transaction - from negotiation to payment to delivery scheduling - happens autonomously between agents, with humans only setting high-level parameters and handling exceptions.

As agents become autonomous economic actors, a significant amount of global GDP will follow one of these two patterns.

Transactions as a reward function

The emergence of agent-driven transactions could help solve one of the thorniest problems in AI: defining reliable reward functions for training agent behavior. In most AI systems today, we struggle to specify exactly what 'good' performance looks like. But commercial transactions have a built-in signal of success: willing payment.

When an agent successfully completes a transaction (whether with a human or another agent), the counterparty's willingness to pay serves as a clear indicator of satisfaction. This creates a natural feedback loop: agents who consistently complete satisfactory transactions get more business, while those who don't will be selected out of the market. Over time, this market-based selection pressure could drive agents to develop more sophisticated and reliable behaviors.

This mirrors how human markets naturally select competent service providers. Commercial activity could become not just an application of AI agents but a crucial training ground for developing more capable and reliable AI systems. The market itself could become the reward function.