Sharded agent simulations: workers, checkpoints and faults

Separate the evaluator from the agents being measured

The sharded simulation reference describes bounded workers, evaluator-owned checkpoints and cross-shard delivery. The evaluator's record matters because the system under test should not define success merely by reporting that it finished.

For an illustrative coordination campaign, establish which observations count toward completion and which artifacts the evaluator must retain. An agent's final message is one input to analysis, not a substitute for the campaign's declared evidence.

Register how work and faults are distributed

The evaluation methodology calls for scenario, topology, population, seeds, software revision and fault profile before execution. Add the shard arrangement relevant to your run to that registration.

If a fault crosses a shard boundary, preserve the intended delivery and observed result. Changing partitioning or fault timing after seeing a failure changes the experiment. Keep exploratory adjustments distinguishable from the registered comparison.

Checkpoint evaluator progress as evidence

A campaign can outlast one worker process. Evaluator-owned checkpoints help identify retained progress without letting a worker's local memory become the only record of completed evaluation.

Design your run review to detect missing or mismatched artifacts after a restart. This is a recommended operational check, not a claim that every interruption or external model request is automatically recoverable. Use the selected simulation implementation's actual restart contract.

Preserve attribution across shard outputs

Immutable, independently attributable artifacts make it possible to connect a result to the worker, scenario and version that produced it. Keep raw shard outputs separate from an aggregate report.

The reproducibility guidance explicitly separates registrations, raw evidence, evaluator checkpoints and interpretation. A merged success count without the contributing records cannot show whether a shard was omitted or a failed attempt was replaced.

Report scope alongside the aggregate

• State the population and shard arrangement actually executed.
• Identify incomplete workers and unresolved deliveries.
• Retain the configured fault schedule and observed deviations.
• Report resource costs as well as useful outcomes.
• Limit conclusions to the measured campaign.

The limitations reference distinguishes implementation checks from universal scale or production reliability. More simulated agents expand the tested scenario; they do not by themselves establish a deployment guarantee.