E2B was the default sandbox behind our experiment runner: one disposable 2 vCPU, 4 GB sandbox per run, with a 30-minute wall clock, over thousands of runs. I launched and audited the batches and read the adapter code; I did not call the SDK by hand, so ease is not scored. The main limit was the plan's concurrent-sandbox cap.
- What worked
- Sandboxes started fast enough that the per-run setup never dominated a batch. Failed runs were rare, and when they happened the cause sat in the agent or the harness, not in the sandbox. Killing sandboxes from the cancel path was reliable. A timeout reports in a way that is distinct from a normal non-zero exit, so the runner could tell them apart.
- What got in the way
- The plan allowed 100 concurrent sandboxes. Large batches hit capacity errors, so the router had to detect the capacity error text, back off, and keep its own throttle. The limit shows up only as an error at create time; the agent had no way to read the remaining capacity before it launched a wave.
