Beyond Sandboxes: Architecting Durable Runtimes for AI Agents
The agent infrastructure conversation has converged on sandboxing: how an agent’s tools execute safely. But sandboxing solves tool execution, not agent operation. The agent itself — the loop making decisions, calling tools, waiting on humans, recovering from crashes, and coordinating work — often still runs in memory. When that process dies, the agent loses its operational state. A sandbox does not help you replay a decision, trace a tool call, reconstruct context, or explain an outcome after the fact.
This talk is about the layer underneath the agent: the runtime it runs on, distinct from the sandbox its tools run in. We will walk through the architectural decisions that determine whether an agent system can:
Replay prior decisions without re-prompting the model
Treat LLM calls, tool calls, state transitions, and human inputs as first-class execution records
Reconstruct why an agent took a path hours, days, or months later
Execute deterministic workflow graphs as part of a broader agent loop
Recover from crashes, retries, and long-running waits without losing state
We will show how agents built with frameworks such as LangGraph, Google ADK, Vercel AI SDK, and OpenAI Agents SDK can run on a durable execution substrate without rewriting the agent logic. The examples use Conductor and Agentspan, but the patterns apply broadly to production agent infrastructure.
You will leave with a practical mental model for what an agent runtime must provide that a sandbox cannot: replayability, traceability, recovery, and operational explainability.
