Electric SQL launches Electric Agents, challenging the VM sandbox status quo with a stream-backed, serverless architecture for durable agent workflows.

While most platforms default to running agents in persistent, stateful VMs (like E2B or OpenAI's Code Interpreter) for security and tool runtime purposes, Electric SQL makes a compelling case that this is an architectural dead-end for scaling long-lived production workflows. Shifting agent durability to the data layer (the append-only stream) and compute to stateless serverless functions represents a massive leap in efficiency, observability, and simplicity, though developers running untrusted code will still need some secure execution environment.

* Data-centric over compute-centric: Modeling agents as persistent data entities (rather than active running processes) allows them to sleep and scale to zero, significantly lowering infrastructure costs.

* Local-first integration: By running the agent SDK within the developer's process, companies keep total control over secrets and backend systems without exposing them to third-party hosted sandboxes.

* Composability: Built-in primitives allow teams of agents to coordinate via streams, enabling complex manager-worker or pipeline architectures.

* Sandboxing caveats: While stateless logic is great for system integration, agents that must execute untrusted code (e.g., executing arbitrary Python scripts) will still require traditional VM sandboxes.