Pay-per-inference for AI agents: How BlockRun and Incarna use Amazon Bedrock AgentCore payments

Overview of Amazon Bedrock AgentCore Payments

Amazon Bedrock AgentCore payments is a managed service designed to enable AI agents to make on-demand payments for services they use during execution. This capability is particularly useful in scenarios where agents need to pay for model inference, API calls, or other microtransactions frequently and in small amounts. The system is built to handle the complexities of such transactions, including payment signing, spending limits, and integration with external services.

Integration with BlockRun and Incarna

BlockRun, a pay-as-you-go inference router, and Incarna, a platform that provides persistent identities for AI agents, have leveraged Amazon Bedrock AgentCore payments to streamline their workflows. BlockRun serves model inference from multiple providers, and each call is quoted, paid, and settled independently. Incarna, on the other hand, ensures that each agent has a unique identity with its own wallet, enabling secure and attributable transactions.

Key Features of AgentCore Payments

AgentCore payments offers several features that make it well-suited for pay-per-inference scenarios. These include:

• Managed Wallets: Incarna provisions each agent’s wallet using the Coinbase CDP connector, ensuring that the customer retains ownership and grants Incarna the necessary permissions.
• Native Protocol Handling: When a service returns an HTTP 402 status code, the agent automatically initiates a payment over x402, signing the transaction with the configured wallet and returning cryptographic proof to the merchant.
• Spending Governance: Spending limits are enforced at the infrastructure level, preventing agents from exceeding predefined budgets even if their prompts are manipulated.
• Settlement Auditing: Payments are settled in stablecoins, such as USDC on the Base network, and each transaction is verifiable on-chain.

Implementation and Workflow

The integration between BlockRun and Incarna demonstrates how AgentCore payments can be implemented efficiently. The process involves setting up payment credentials, creating a payment manager, and defining spending limits. Once these steps are completed, the agent can initiate a payment session, and AgentCore payments will handle the signing and verification of each transaction.

The workflow for a single inference call is as follows:

1. The agent requests a model inference from BlockRun.
2. BlockRun returns a price quote and initiates a payment session.
3. AgentCore payments checks the quote against the spending limits and signs the transaction using the agent’s wallet.
4. BlockRun verifies the payment and delivers the inference.
5. The transaction is settled on-chain, ensuring transparency and auditability.

Benefits and Use Cases

The use of AgentCore payments provides several benefits, including reduced development time, improved security, and better control over spending. By leveraging managed services, developers can avoid the need to build custom payment infrastructure, which can be time-consuming and error-prone. This approach is particularly useful for applications that require frequent, low-value transactions, such as AI agents interacting with multiple model providers.

Additionally, the system supports two x402 payment schemes: exact and upto. The exact scheme is used when the price is known in advance, while the upto scheme allows for dynamic pricing with a ceiling limit. This flexibility ensures that agents can operate efficiently in a variety of scenarios.

Conclusion

Amazon Bedrock AgentCore payments provides a robust and scalable solution for enabling AI agents to make on-demand payments. The integration with BlockRun and Incarna demonstrates how this system can be used to streamline pay-per-inference workflows, reduce development time, and ensure secure and auditable transactions. As AI agents become more autonomous and complex, the ability to manage payments at scale will become increasingly important.

Source: AWS Machine Learning. Analysis based on publicly available information; no hands-on testing is claimed.