Agentic RL infrastructure · source snapshots 17, 21 and 22 September 2026

Agentic RL infrastructure, as the reports describe it

Agentic RL is where production stacks publish the most infrastructure detail and public benchmarks exercise the least. This page collects that detail: sandbox and harness fleets, staleness knobs and in-flight handling, a failure ledger, and the mechanisms that only exist across devices; the loop itself is on its own page.

Sandbox and harness fleets

Environment execution is a first-class service scaled separately from GPUs. The concurrency figures are first-party claims, none independently verified.

Staleness knobs and in-flight handling

Every 2026 system runs bounded-staleness asynchrony, but each counts staleness differently, so values are not comparable without the definition (derivations in ASYNC_RL_STALENESS.md).

Failure ledger

What reports admit went wrong and how the run recovered. Pretraining entries are included because the same hardware and scheduler failures hit RL clusters.

Mechanisms that only exist across devices

On one GPU there is nothing to place, synchronize or recover; these seven mechanisms appear only once a run spans two or more devices.

Which audited benchmark tasks reach agentic RL

Every entry is related evidence: fixed-model harnesses, an application patch, an agent-memory design. None trains a policy through tools or exercises the loop's infrastructure.

Sources on this page

Show sources