Quick Overview
This interview clip from How I AI features Sharadh Krishnamurthy, an Engineering Manager at Stripe, discussing the infrastructure challenges of deploying AI agents. Krishnamurthy shares early operational lessons and explains how Stripe approaches agent identity and infrastructure reliability.
Key Points
- 1.Stripe is exploring agentic identity to determine what an AI agent is and what specific use case it is attempting to execute across internal infrastructure.
- 2.Establishing agentic identity helps engineering teams manage infrastructure priorities and execute load shedding effectively.
- 3.Deploying AI agents amplifies conventional infrastructure failure modes and increases the amplitude of operational problems.
- 4.Early internal agent deployments at Stripe experienced rogue agents that nearly took down core infrastructure systems.
- 5.Stripe caught the rogue agent issues in time and has hardened its core infrastructure systems against agent-induced disruptions.
Summary
Sharadh Krishnamurthy, an engineering manager at Stripe, explains how the company is approaching the concept of agentic identity within its infrastructure. While acknowledging that the problem is not yet fully solved, he describes the need to identify specific AI agents and clarify their precise use cases as they interact with internal infrastructure. Establishing this visibility allows engineering teams to apply foundational infrastructure practices, such as managing request priorities and performing load shedding.
Krishnamurthy notes that running AI agents inside an infrastructure environment significantly increases operational risk. Rather than introducing entirely unfamiliar failure types, agents dial up existing failure modes and multiply the amplitude of system problems.
During initial implementations, Stripe encountered situations where agents went rogue and came close to taking down core infrastructure systems. The engineering team caught these incidents in time to avoid outages and subsequently hardened their core systems against similar agent-driven failures.
Defining Agentic Identity in Infrastructure
Sharadh Krishnamurthy discusses how Stripe approaches agentic identity, exploring methods to identify specific AI agents and their intended use cases across the infrastructure to govern traffic priorities and load shedding.
Amplification of System Failure Modes
Operating AI agents in production environments magnifies traditional infrastructure risks by dialing up existing failure modes and increasing the scale of potential problems.
Addressing Rogue Agents and System Hardening
Stripe encountered operational challenges where early agents went rogue and nearly disrupted core systems, prompting the engineering team to intervene in time and harden the underlying infrastructure.
The Bottom Line
The video establishes that deploying AI agents exacerbates existing infrastructure vulnerabilities, making identity tracking and load management essential. While Stripe prevented outages from early rogue agents and hardened its core systems, developing a complete solution for agentic identity remains an ongoing effort.
FAQ
What is agentic identity when managing AI agents within Stripe infrastructure systems?
Agentic identity is the practice of identifying an AI agent and determining its specific intended use case as it operates across internal infrastructure. This identity is used to manage operational priorities and implement load shedding.
How do AI agents impact system failure modes and infrastructure stability at Stripe?
AI agents dial up existing system failure modes and multiply the amplitude of operational problems across infrastructure.
What specific operational challenges did Stripe experience when running early internal AI agents?
Stripe experienced incidents where early AI agents went rogue and nearly brought down core systems before engineers caught the issues in time and hardened the infrastructure.
Worth watching for
Infrastructure engineers and engineering leaders managing AI agent deployments who want to understand the reliability risks and operational governance required for autonomous agents.
- ai-agents
- stripe
- infrastructure
- system-reliability
- engineering