Agent environments
Real work, made reproducible.
We reconstruct consequential workflows as controlled environments for evaluation, reinforcement learning, and system development.
Task
Representative work, inputs, constraints
Environment
Tools, permissions, workspace state
Trajectory
Messages, actions, recovery, provenance
Artifact
The document, workbook, deck, or decision
Grader
Rubric evidence and expert judgment
Verdict
Comparable result and deployment signal
Environment anatomy
Enough reality to expose failure. Enough control to learn from it.
Workspace
The files, applications, state, and context the agent actually receives.
Tools
The same operations required by the work—not simplified proxy actions.
Observability
Messages, tool calls, intermediate files, recovery, latency, and cost.
Judgment
Artifact-aware graders, deterministic checks, and expert rubric evidence.
Open evidence
See the environment through the run.
A published trace is not a product screenshot. It is the prompt, workspace, execution, resulting artifact, and evaluation record visitors can inspect themselves.
Review warranty claims and update refund amounts
Review the attached warranty claims, then edit the existing product purchases spreadsheet to show the maximum refund amount a customer could receive for each product purchased.
Failure mode · Missed governing agreement constraints
GPT-5.5 · xhigh · Raycaster harness · public tier
Live public run · transcript · files · rubric
Inspect full run ↗From one workflow to a program