How Agentforce Runs Secure AI Agents at 11 Million Calls Per Day

Summary

The article provides an in-depth look at how Salesforce's Agentforce platform manages over 11 million daily agent calls while ensuring security and reliability. It highlights the architectural design of a layered access-control plane that enforces identity, authorization, and request integrity across a multi-tenant environment. The article emphasizes the importance of blast-radius containment and the separation of responsibilities within the access layer to mitigate risks associated with unauthorized access and service degradation. Additionally, it discusses the necessity of observability in managing transient states during agent interactions, ensuring that the platform can evolve rapidly without compromising security.

Key Learnings

• 1The access-control plane is designed to enforce strict authentication and authorization, ensuring that only valid requests reach backend services.
• 2Separation of responsibilities across platform boundaries reduces operational risk and avoids duplication of security controls.
• 3Blast-radius containment is critical in a multi-tenant environment to prevent issues from affecting multiple customers.
• 4Session-based tracking is implemented to enhance observability and reliability analysis of agent interactions.
• 5Continuous automated monitoring is essential for maintaining performance and availability at high request volumes.

Who Should Read This

Senior Software Engineers specializing in security architecture and scalability in multi-tenant environments.

Test Your Knowledge

Topics