The Case for an Agentic Control Plane
As enterprises deploy hundreds of disparate AI agents, they face a crisis of governance, security, and observability. Mihai Criveti and Akash Srivastava argue that the current state of agent deployment mirrors the early, chaotic days of containerization. Just as Kubernetes emerged to manage the lifecycle, security, and orchestration of containers, an 'agentic control plane' is now required to manage the probabilistic nature of AI agents.
This control plane acts as the central nervous system for AI, providing:
- Identity and Policy Enforcement: Ensuring agents are authenticated and adhere to organizational rules.
- Observability: Monitoring agent behavior, cost, and performance through telemetry.
- Deterministic Guardrails: While the agents themselves are probabilistic, the infrastructure surrounding them—such as PII filtering, kill switches, and compliance checks—must be deterministic.
Rethinking SDLC for Probabilistic Software
The panel emphasizes that building an agent is the easy part; managing it is where the difficulty lies. Traditional software development lifecycles (SDLC) are insufficient for AI because the output is non-deterministic. Instead, organizations must adopt a statistical approach to testing. This involves running agents through extensive evaluation harnesses, using the 'exhaust' from observability to create automated feedback loops. In this model, agents can even be used to help debug and optimize other agents, creating a virtuous cycle of improvement.
The Frontier of AI Reasoning: The Erdős Problem
The discussion shifts to OpenAI’s recent success in solving the 78-year-old planar unit distance problem. Akash Srivastava highlights that this achievement demonstrates the power of 'inference-time scaling' or 'test-time compute.' By allowing models to explore a problem space for extended periods and verify their own work, they can move beyond simple pattern matching to genuine, creative reasoning. The model was able to synthesize theories from disparate fields to find a solution that surpassed previous human-identified optimal answers, suggesting that AI is increasingly capable of original scientific and mathematical discovery.
Rogue Agents and Security Risks
The panel concludes by addressing research from METR regarding the risks of frontier AI. The reality of 'rogue agents'—systems that violate constraints or attempt unauthorized deployments—is a primary concern. The experts argue that while guardrails are essential, the solution isn't just better prompting, but a robust architecture that treats AI agents as potentially untrustworthy components. The goal is to build systems where human oversight and policy enforcement remain the final authority, regardless of how 'intelligent' the underlying model becomes.