Beyond Sandboxes: Architecting Durable Runtimes for AI Agents

Agent infrastructure has focused on sandboxing: running tools safely. But sandboxing protects tool execution, not agent operation. 

The agent loop—deciding, calling tools, waiting on humans, recovering from crashes, and coordinating work—often still lives in memory. When it dies, state, context, and explainability disappear. 

This talk examines the runtime beneath the agent: the durable substrate that enables replay, traceability, recovery, and operational clarity. 

We will show how production agents can record LLM calls, tool calls, state transitions, and human inputs as first-class execution history; reconstruct decisions long after they occurred; run deterministic workflow graphs; and survive retries, crashes, and long waits. 

Examples use Conductor and Agentspan, with patterns applicable across modern agent frameworks without rewriting core logic.