Thesis: For agent sandbox infrastructure, useful CPU comparisons must measure verified task throughput under realistic concurrency, not just single-task speed.
Observed facts: Modal says its Agentless Agent Bench contains 50 Terminal-Bench oracle solutions and verifiers with network calls removed, run in concurrent, prewarmed Docker containers. Modal reports that Vera often had only modest or inconsistent advantages in serial tests on an unloaded machine, while its throughput test showed roughly twice the task completions under heavy load without a corresponding latency penalty. Mechanism: Concurrent CPU-bound bursts compete for host resources; if one CPU sustains more completed tasks as concurrency rises, the scheduler may obtain more useful work from a host without simply packing it more densely and transferring the cost into user-visible delay.
Why now: Modal is evaluating Vera as it extends an historically x86 infrastructure stack to Arm, and presents the result as relevant to sandbox users and to workflows where CPU time gates downstream GPU work. First implication: Teams choosing compute for agent sandboxes should benchmark their own mixed concurrent tasks and latency objectives rather than infer value from isolated CPU specifications. Evidence boundary: These are Modal’s internally designed workloads and vendor-reported results; the article does not establish independent replication or performance across all sandbox workloads. First watch condition: Watch for independent replications or production measurements showing whether the throughput and latency pattern persists across other agent workloads and host configurations.