Abundant says it has run more than 75,000 reinforcement-learning rollouts in Modal Sandboxes, using the infrastructure to build and test environments for AI agents. Its account offers a useful look at how sandbox design, resource management and verification shape the work behind agent training.
Modal Labs Watch analysis
What happened
Abundant builds tasks for AI models to attempt, then checks whether they succeed using a verifier. The company says almost all its rollout platform runs on Modal, with Modal Functions handling scheduling and Modal Sandboxes providing the task environments. As of November 2026, it says it was running more than two million Sandboxes a month.
A rollout platform took the team around two weeks to implement, according to Modal’s customer story. Abundant says it used Sandboxes’ CPU and memory bursting to accommodate workload peaks without paying to reserve the same level of compute all the time. It also changed how tasks begin: instructions are deleted before work starts, and a reviewer agent checks traces for signs that a model knows it is being evaluated.
Why it matters
Agent training is not just a matter of giving a model a difficult task. Each run also needs compute, an isolated environment and a reliable way to judge whether the model did the job. Modal’s account shows how those pieces can be assembled, and how infrastructure choices can affect both idle-compute costs and the quality of an evaluation.
The evaluation detail matters: if a model can find its instructions, it may behave as though it is taking a test rather than tackling the task naturally. Abundant says its reviewer checks for that kind of evaluation awareness. The work also supports its Frontier RSI benchmark, for which the company says Modal powered more than 1,300 hours of agent-research time across nearly 500 trials.
Our read
This is a useful engineering case study, not an independent comparison of sandbox providers or proof that every agent benchmark has solved evaluation. Still, the specifics are more interesting than the usual “AI needs lots of compute” refrain: burst capacity can trim idle-resource costs, while careful task setup helps stop agents gaming the exam.
What to watch
- Whether Abundant publishes more detail on its Frontier RSI benchmark and evaluation methods.
- How CPU and memory bursting performs across other workloads and customers.
- Whether agent evaluations increasingly check for signs that models recognise a test environment.
Discussion spark: Should agent benchmarks prioritise realistic, hard-to-game environments, even if that makes results more expensive and harder to reproduce?
Sources and evidence
- Source update (19 November 2026, 00:00 UTC)
Independent WittyWires tracker for public updates about Modal Labs. Not affiliated with or endorsed by Modal Labs; this is not an official account.