What problem does it solve? RL training rollouts stall when sandbox capacity is mis-sized: pools built for the wrong concurrency queue claims, autoscaling ceilings get mistaken for requests, and quota limits surface only mid-run. This Skill walks through sizing a warm pool to the trainer's actual peak concurrency and driving sandboxes correctly so episodes start in about a second instead of a minute. ## Core Features & Use Cases - Concurrency-first pool sizing: Asks for peak concurrent sandboxes (not total episodes), then sets replica targets, minReplicas floors, and autoscaling ceilings with abx scale and scaling-group updates. - E2B SDK rollout loop: Provisions capacity once with abx, then claims, runs, and discards sandboxes per episode through the standard E2B SDK, with env docs providing the API URL, domain, and scheme. - Stall diagnosis: Covers the three failure modes in order — pools stuck below demand, quota ceilings, and sandbox runtimes failing to start — with the command to check each. - Use Case: A trainer needs 10,000 episodes at 64 concurrent environments. Scale the pool to 64, set minReplicas to the steady-state floor, point the E2B SDK at the env's documented endpoint, and tag each sandbox so runs stay distinguishable. ## Quick Start Ask the assistant to set up an AgentBox warm pool sized for your peak concurrency and write an E2B SDK rollout loop that claims a sandbox per episode.
