Making Expensive Agentic Research Survive Reality

Praxist's exploration mechanisms only matter if the resulting work can run, be paid for, and be stopped without losing the evidence that made it useful. This chapter examines the execution substrate beneath the research loop—resource ownership, dynamic budgets, runtime normalization, prompt layout, and cost signals—so a reader can tell the difference between a healthy-looking process and a scientifically trustworthy result.

Resource ownership is scientific provenance

A research experiment has two kinds of identity. It has a scientific identity—what variant, protocol, data coverage, and hypothesis are being tested—and an operational identity—what process may start, which resources it receives, and when those resources are released. If different components own those facts, the run can lose track of whether two jobs are duplicates, whether they used the intended device, or whether a result belongs to the experiment that requested it.

When central scheduling is selected, Praxist gives one scheduler responsibility for four linked facts: which scientific experiment was submitted, which attempt may start, which process group and physical accelerator were assigned, and when the allocation is released. Peers submit task-defined work with a stable semantic tag, a resource profile, an estimated duration, and a work class. They do not pick a device or create a scheduled task process on their own.

This arrangement is not merely an infrastructure preference. A resource assignment can affect what a result means. A task that was calibrated for one accelerator profile may behave differently under another. A retry that silently becomes a second independent run can distort the apparent amount of evidence. A process that outlives the agent that launched it can keep consuming a resource after the research loop believes the work is finished. One owner can correlate the scientific request with the process group and the eventual release.

The scheduler does not invent the profile. Task initialization observes the unchanged public baseline and uses that observation to plan concurrency and resource demand. It should not rewrite a GPU task as a CPU task to make arithmetic look convenient, or treat a teardown sample as proof that the task has no accelerator demand. If demand is ambiguous or undersampled, the conservative choice is to keep the assignment exclusive rather than claim precision the measurement does not support.

Natural-unit parallelism is another guard against false scale. A task should expose independent complete-evaluation units—such as seeds, folds, scenarios, simulator instances, or benchmark cases—only when those units can be distributed, retried, and aggregated without changing the protocol. Declaring several GPUs does not create parallelism if the evaluator serially performs one unit or fails to preserve device binding through its descendants. Conversely, counting internal workers as separate top-level experiments can double-count the same scientific work.

Semantic tags make retries safer. The tag identifies the scientific experiment, not its timestamp, output path, or retry number. Repeated submissions with the same identity can share one queued, running, or completed job. An infrastructure failure may create another attempt under the same semantic identity, while a changed scientific request must not masquerade as a retry by appending arbitrary text. This is how the scheduler can distinguish “the same experiment recovered” from “a new experiment happened to use a similar command.”