Sandboxes & runtimes

Isolated sandboxes an agent can start in milliseconds to run the code it wrote.

The sandbox layer underneath every coding agent — the place where "give the model its own computer" stops being a metaphor and becomes a per-second price.

A fresh microVM per agent session, so generated code runs away from your servers.

The sandbox layer — whoever owns "where the agent's code actually runs" sits under every coding agent and every computer-use product, and E2B is the one selling it as a primitive rather than as part of a platform.

Durable workflows — functions that survive crashes, retries and deploys.

The durable-execution layer that agent builders reach for when "retry the whole thing" stops being an acceptable answer — and one of the few in that layer publishing its own reliability post-mortems.

A headless browser built for machines, starting in milliseconds on little memory.

They are rebuilding the browser on the premise that nobody is looking at it — which is the cheapest, most measurable claim on the roster, and nobody has independently benchmarked it against headless Chrome.

Journals every step of a service so it resumes exactly where it died.

The durable-execution argument at its most technical — a Flink-lineage team insisting that an agent loop needs a log underneath it, not a checkpoint file, and publishing the reasoning rather than the marketing.

Runs long background jobs and agent loops in TypeScript with no timeout.

The durable-execution layer that agent builders reach for first because it is just TypeScript — the clearest test of whether "make the agent loop survive a restart" becomes a product category or a framework feature.