Preparing your design workspace

Loading the problem and your latest saved attempt…

Design a Long-Horizon Agent Training Environment Platform

Run reproducible long-horizon agent trajectories in versioned environments with auditable rewards, checkpoints, evaluation, and safe promotion.

Problem statement

For this exercise, assume a platform for training and evaluating agents on long-horizon tool-use tasks in a sandboxed versioned environment.

A TaskDefinition fixes the initial observation, permitted tool actions, environment image, hidden success criteria, maximum 1,000 steps, and terminal success, failure, timeout, or invalid-action outcomes.

Assume a trajectory uses stable episode and step identities, records observations, actions, tool results, checkpoints, and environment state hashes, and links one immutable RewardDefinition.

Rewards combine deterministic task evidence with separately versioned reviewer judgments; training optimization itself is outside scope.

The platform schedules many sandbox rollouts, resumes only from validated checkpoints, detects reward gaming and environment drift, compares policies on immutable evaluation suites, and promotes through guarded release decisions.

Design task and environment publication, rollout execution, checkpointing, reward computation, review, evaluation, isolation, privacy, recovery, and audit. Building the learning algorithm or production agent application is outside scope.

What to build

  1. Publish immutable compatible task, environment, action, terminal, and reward definitions.
  2. Execute one isolated retry-safe long-horizon trajectory with bounded steps and validated checkpoints.
  3. Compute attributable deterministic and reviewed reward evidence without rewriting prior outcomes.
  4. Compare policy versions on immutable suites and record one guarded promotion or rollback decision.

At scale

  • For this exercise, assume 100,000 rollout starts per day, 10,000 concurrent sandboxes, and trajectories of up to 1,000 steps or 24 hours.
  • For this exercise, persist accepted step evidence within five seconds and create a validated checkpoint at least every 50 steps or 15 minutes.
  • For this exercise, retain trajectory, reward, evaluation, and release metadata for one year while task data follows its owning access and deletion policy.
  • For this exercise, process or zone failure must not duplicate a step or reward contribution; an unrecoverable environment occurrence ends explicitly rather than silently restarting as the same episode.