Agent environments

Real work, made reproducible.

We reconstruct consequential workflows as controlled environments for evaluation, reinforcement learning, and system development.

01

Task

Representative work, inputs, constraints

02

Environment

Tools, permissions, workspace state

03

Trajectory

Messages, actions, recovery, provenance

04

Artifact

The document, workbook, deck, or decision

05

Grader

Rubric evidence and expert judgment

06

Verdict

Comparable result and deployment signal

Environment anatomy

Enough reality to expose failure. Enough control to learn from it.

01

Workspace

The files, applications, state, and context the agent actually receives.

02

Tools

The same operations required by the work—not simplified proxy actions.

03

Observability

Messages, tool calls, intermediate files, recovery, latency, and cost.

04

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.

Published runs · live product surface
APEX-Agents · Law5 / 5 · Pass6m 12s

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

Build evaluations, training episodes, and regression tests from the same environment.