AI agents increasingly need to make frequent, small payments for services like model inference, API responses, or web content access. These micro-transactions, often less than a cent each, occur autonomously within the agent’s operational loop, without human oversight. Managing these payments traditionally presents challenges, including supporting sub-cent payments on conventional card networks, implementing payment protocols, securing funds, and preventing agents from overspending.
Amazon Bedrock AgentCore Payments addresses these challenges by providing a managed capability that allows AI agents to pay for services on demand. It handles the payment protocol, connects to a digital wallet, signs transactions, and enforces spending limits at the infrastructure layer, making it suitable for high-frequency, low-value payment patterns like pay-per-inference.
How AgentCore Payments Works
AgentCore Payments integrates with a customer's wallet, such as one provisioned through the Coinbase CDP connector. The customer retains ownership of the wallet and grants delegated authorization for the agent to use it. When an agent needs a paid service, like model inference from a provider such as BlockRun, it initiates a request. If the endpoint responds with an HTTP 402 “Payment Required” challenge, the agent uses AgentCore Payments to process the payment over the x402 protocol. This involves signing the transaction with the configured wallet and providing cryptographic proof to the merchant.
A key feature is the enforcement of spending limits at the infrastructure layer, ensuring that an agent cannot exceed its allocated budget, even if its prompt is manipulated. Payments are settled in stablecoins, such as USDC on the Base network, and each transaction is verifiable on-chain, providing an auditable record. This architecture allows for a pay-per-inference model where agents only pay for what they use, and unmade calls incur no cost.
Example Flow: Buying a Single Inference
Consider an agent requiring a model call. It integrates with BlockRun, an inference router that serves over 90 models from more than 15 providers. BlockRun responds with a PaymentRequired challenge, quoting a price for that specific call. The agent then opens a payment session and calls ProcessPayment via AgentCore Payments. This service checks the quote against predefined spending limits and signs the authorization from the agent's wallet. Once the seller (BlockRun) verifies the payment signature, it serves the inference and records the charge, which is a small per-call amount.
AgentCore Payments supports two x402 payment schemes: 'exact,' typically used when the price is known upfront, and 'upto,' suitable for dynamically priced resources. With 'upto,' the agent authorizes a maximum amount, and the inference provider settles for the actual usage, up to that ceiling.
Controlling Agent Spend
To provide comfort in letting agents handle real money, AgentCore Payments uses payment sessions. Each session sets a maximum spending limit and an expiry time, enforced by the infrastructure. For instance, Incarna, a user of AgentCore Payments, sizes its sessions to a day’s budget. This prevents an agent from overspending, even if its internal logic goes awry. This session-level budgeting can be adopted without code changes, even for existing integrations.
Practical Uses and Limitations
The implementation of AgentCore Payments significantly reduces the development effort for integrating pay-per-inference capabilities. Incarna reported completing their full integration in three days, with approximately 200 lines of application code, a substantial reduction from an original estimate of two to three months. During beta testing, agents processed over 1,000 payments, ranging from $0.001 to $0.05 per call, all settled individually on-chain.
This capability is crucial for scaling AI agents into production workflows where economic sustainability and precise cost attribution are necessary. It supports the trend of AI agents moving into tasks that require reasoning over enterprise knowledge, operating across long-horizon tasks, and acting in systems where security, accuracy, and governance are paramount.
BlockRun's founder, Vicky Fu, highlighted that this approach allows developers full control over their model set while benefiting from benchmark-driven routing, leading to higher task success rates at lower token costs. With AgentCore Payments providing spending controls, this intelligence can be safely deployed in real-world scenarios. While the service significantly streamlines payment processes for agents, continuous monitoring and robust governance remain essential for any production AI system.