APEXY LAB / DATA CATALOG

Data that gives agents
somewhere to go.

We turn real work into structured environments, trajectories, and evaluations that an agent can practice and a team can inspect.

Talk about your data Open a sample
TRAJECTORY / 001 VERIFIED SHAPE
01task
02actions
03outcome
state · action · reward · artifactmany → one
01 / WHAT WE SHIP

Built around the work
models need to learn.

Each product starts with a real workflow and ends with an outcome that can be replayed, checked, and improved.

IN DEVELOPMENT
WEB / LONG-HORIZON

Browser workflows

Multi-step tasks with state, tools, recovery paths, and a verifiable end state.

  • Task state
  • Tool calls
  • Final artifact
IN DEVELOPMENT
CODE / TOOL USE

Repository work

Change-making workflows that connect intent, files, commands, tests, and reviewable diffs.

  • Plan
  • Action trace
  • Test result
IN DEVELOPMENT
REASONING / EXPERT

Research tasks

Open-ended work with source material, intermediate decisions, and a rubric for the outcome.

  • Sources
  • Decision log
  • Rubric
IN DEVELOPMENT
EMBODIED / SIMULATION

World interaction

Composable tasks for systems that learn by acting, observing, correcting, and trying again.

  • Observation
  • Action
  • Outcome
02 / ILLUSTRATIVE RECORDS

One record keeps
the whole path.

These examples show the shape of an Apexy trajectory. Production collections are built around each team's task distribution and evaluation needs.

TaskDomainTraceCheck
Browser checkout recoveryWebState → actions → artifact Replayable
Repository migrationCodePlan → diff → tests Diff checked
Multi-source research briefReasoningSources → decisions → rubric Rubric scored
Pick-and-place recoveryRoboticsObservation → action → outcome State logged
03 / THE DATA LOOP

From lived work
to useful signal.

Structure is what makes a trajectory more than a recording. It lets teams see where an agent acted, why it failed, and what to try next.

01

Capture the work

Start with a task that has context, constraints, and a meaningful outcome.

02

Compose the world

Add the tools, state transitions, and dependencies an agent must navigate.

03

Verify the result

Make the target, trajectory, and evaluation criteria inspectable.

04

Feed the loop

Turn what the agent did into the next environment, task, or test.

APEXY LAB / NEXT PATH

Show us the work
you want models to learn.

Start a conversation