Skip to content

Govern the runtime. Not the agent.

with Sundar Vedaraman

Wednesday 25 November 13:30 – 14:00 Stage 3 - powered by GitHub

About This Session

The industry is converging on a multi agent, multi model world. The best model changes every couple of months, it makes no economic sense to call a frontier model to classify a ticket, and unique workflows increasingly need custom models. So teams end up running many models across many harnesses (Claude Code, Copilot, Cursor, custom). That makes trust the defining challenge. Harnesses ship their own guardrails, but each one only governs its own traffic. Agents can route around rails they can't see, and those rails can shift whenever a model vendor ships an update. Safety shouldn't depend on the agent making the right call, or on someone else's release schedule. We'll cover: - What's pushing engineering teams toward running many models across many harnesses, and why this convenience quietly turns into a trust problem - The case for a meta runtime: one layer below every harness, governing execution, tool calls, credentials, scoped permissions and token spend, no matter the model or harness - How a fine-grained, default-deny policy lets agents run at full speed inside the lines, with the audit trail written automatically

Topics

  • AGI (Artificial General Intelligence)
  • AI Coding Assistants
  • AI Models
  • Agents
  • Agentic AI
  • Compliance
  • DevOps
  • Governance