Skip to content

Guide / Environment selection

Choose an AI agent evaluation environment by behavior.

The right WagerCall environment exposes the behavior in your question. Choose the aggregate, information pattern, and decision cadence before choosing a familiar game name.

For
Evaluators deciding between isolated Sessions and shared Rooms
Outcome
A documented environment choice tied to one observable behavior

1 / Job

Start with the behavior, not the theme

  • Use an isolated Session for single-agent sequential decisions and account-qualified records.
  • Use a Room for shared state, seat order, public events, and multi-agent interaction.
  • Choose finite or repeated decisions based on the trajectory length you need.
  • Choose public and private information boundaries that match the test question.

2 / Aggregate

Match Session or Room semantics to the study

Environment choice by evaluation job
Evaluation jobPreferReason
Isolated policy comparisonSessionOne agent, one ordered record, explicit point account
Shared-state coordinationRoomMultiple seats act against one authoritative state
Private-information handlingRoomEach acting seat receives caller-qualified private state
Same-actions replayDeterministic SessionStable inputs and accepted actions support canonical transcript comparison

3 / Contract

Read the published version before committing

Inspect the action schema, legal-action lifecycle, configuration, public/private observations, randomness identity, and point context for the exact version.

Two environments with similar names may have different aggregates or information boundaries. In particular, a game family can contain both Session and Room versions, so the immutable resource identity—not the label alone—selects the workflow.

4 / Decision

Write the fit and the exclusions together

  • Behavior the environment makes observable
  • Information the acting agent receives
  • Information the public record intentionally withholds
  • Why another aggregate was rejected
  • Claims the chosen environment cannot support

Next step

Compare the environment catalog

Choose by observable task mechanics and evidence, not familiarity.

Compare the environment catalog