Labelbox

What is Labelbox?

Labelbox provides AI training environments and autonomous agent workflows for enterprises and frontier AI labs. Its Recursion product runs background agents that handle engineering, research, and operational tasks triggered by alerts, schedules, or API calls, while its Horizon, Terra, and Alignerr products supply post-training data and expert evaluations. Enterprise access requires submitting a specific job or contacting the team to begin setup.

What are the features of Labelbox?

• Autonomous multi-agent execution (Recursion)

  Starts specialist agents based on schedules, alerts, commits, or API calls to complete jobs in parallel within isolated sandboxes.
• Graded session memory

  Evaluates every agent run against a user-defined rubric, storing notes and skills from graded sessions so subsequent runs start with prior context.
• Specialist model training

  Converts frequently run, graded agent sessions into training environments to fine-tune specialist models that replace general-purpose ones once they outperform them.
• Reinforcement learning environments (Horizon)

  Supplies post-training environments and evaluations for reasoning, tool use, and computer use, including simulated enterprise apps built with WorldSim.
• Robotics data capture (Terra)

  Collects egocentric and teleoperated demonstrations using purpose-built rigs with seven synchronized viewpoints and an AI diversity engine to target missing corpus data.
• Expert evaluation network (Alignerr)

  Provides rubrics, preferences, and graded assessments written by over 2.6 million skill-tested experts across more than 200 domains.

What are the use cases of Labelbox?

• Model regression triage

  A coordinator and five specialist agents rerun evaluations, diff training data, bisect code, and ablate changes to trace a checkpoint score drop, delivering a fix for review.
• AI benchmark creation

  Labelbox produced 820 expert-authored problems and 528 rubric-graded prompts for Meta's GIM benchmark, enabling calibration over more than 200,000 prompt-response pairs.
• Automated dependency upgrades

  Agents detect new package versions, upgrade dependencies, fix resulting breakages, and deliver one ready-to-merge pull request per package with passing tests.

How to use Labelbox?

1. Define the job

   Describe the recurring task, the trigger such as an alert or schedule, the connected applications, and the rubric defining a successful result.
2. Agent orchestration

   Recursion plans the work, provisions sandboxes and credentials from a vault, and assigns specialist agents to run tasks in parallel.
3. Review delivery

   Receive the finished output, such as a pull request, results table, or Slack thread, along with the attached evidence and grading score.
4. Approve or edit

   Approve the delivered work or edit it before the agent applies the result to your systems.