Archon: keeping financial facts outside the LLM, with Strands Agents

Maintaining factual integrity in financial communications is crucial. Archon's "Agents for Humans" project aims to preserve the accuracy of invoices and payments by separating the generation of prose from the verification of financial details. Six readers, each specialized in a particular aspect such as suppliers, sales, or payroll, provide reports that are then combined before a final draft is prepared.

This multi-disciplinary review ensures that the final message is both accurate and appropriate for the intended recipient.

The architecture of Archon is designed to be robust and secure. Python Strands Agent instances are used to handle specific tasks, such as drafting and verification, within a controlled environment. These agents are scoped to perform particular functions, and their outputs are strictly validated before being integrated into the final draft. This separation of concerns ensures that the automated system cannot generate arbitrary tools or send emails without explicit authorization.

One of the key features of Archon is its deterministic ledger, which calculates the eligible balance based on recorded payments and outstanding invoices. This ledger is the authoritative source for any financial claims made in the drafted messages. The system employs Amazon Bedrock for text generation and Amazon SES for email sending, both of which are invoked through a secure API layer.

This API layer does not have direct access to these providers, ensuring that all interactions with external systems are properly authorized and logged.

The approval process in Archon is designed to be immutable. Once an approval is granted, it cannot be revoked unless new evidence changes the circumstances. This approval is tied to specific details such as the recipient, message content, balance, and the workspace revision number. By linking the approval directly to these immutable factors, Archon ensures that any changes to the facts cannot be used to retroactively alter the approved message.

This creates a clear audit trail that can be reviewed at any time to verify the integrity of the communication.

The Archon system is also designed with a clear separation between the API and the provider worker. The API layer, which includes the frontend hosted in a private Amazon S3 bucket, only interacts with the Bedrock and SES services through a Lambda function. This separation prevents unauthorized access and ensures that any errors or inconsistencies in the provider's operations can be easily reconciled without affecting the user's workflow.

The use of S3 for persistence ensures that data is resilient to updates and can be reliably restored if needed.

In demonstrating its capabilities, Archon provides a live demo that includes real-time text generation, human review, and actual email sending through Amazon SES. However, it is important to note that the demo uses synthetic data and does not represent a real financial transaction. The system's purpose is to illustrate how the different components work together to maintain the integrity of financial communications, rather than to provide a production-ready solution.

In conclusion, Archon's architecture is built on the principle of separating duties and responsibilities. By keeping the generation of financial communications distinct from the verification of financial facts, and by ensuring that approvals are tied to immutable evidence, Archon provides a transparent, auditable, and secure way to manage invoices and payments. This approach not only protects the accuracy of financial communications but also provides a clear framework for evaluating and improving the system over time.