Horizon

What it does

Horizon is Labelbox's service for building reinforcement learning (RL) training environments and evaluations for AI models. Inputs include task definitions, domain expertise, verification rules, and reward goals. Outputs include software-generated RL environments with reward signals, preference labels that structure human comparisons of agent trajectories, and WorldSim, a simulator of enterprise tools such as GitLab, Jira, CRM, email, and chat populated with realistic business data.

Teams post-training models on long-horizon work can use Horizon for tasks spanning autonomous research, agent coding and software engineering, multimodal knowledge work, voice with tool use, computer use, and cybersecurity. The distinguishing emphasis is reward design: verification, difficulty progression, credit assignment, and pass@k targets, aimed at keeping models from learning shortcuts. Labelbox cites a Meta project in which its expert-authored data was used to build a benchmark for evaluating frontier model reasoning.

Horizon is aimed at AI labs and teams training or evaluating agentic models that need domain-specific RL data and environments. Engagement starts with a request to contact Labelbox rather than self-serve signup. Buyers should confirm scope, deployment, and how much customers can configure the environments themselves directly with the vendor.

What it can do

• Generates software-based RL environments with calibrated reward signals and pass@k optimization targets
• Simulates enterprise knowledge-work environments (GitLab, Jira, CRM, email, chat) with configurable scenarios
• Produces preference labels that structure human judgments over agent trajectories
• Builds long-horizon tasks with intermediate reward signals across research, coding, multimodal, voice, computer-use, and cybersecurity domains
• Supplies expert-authored problems with structured rubric scoring and quality assurance

Example workflows

Building an RL environment for agent coding

Input: a target domain such as production debugging, reward goals, and pass@k targets. Steps: Labelbox designs long-horizon tasks on real codebases, sets verification and intermediate reward signals, and calibrates difficulty. Output: a software-generated RL environment with calibrated reward signals for post-training.

Collecting preference labels on agent trajectories

Input: agent responses or trajectories for a set of tasks. Steps: human reviewers compare trajectories and record structured preferences. Output: preference labels capturing which long-horizon responses are judged better, for use in model training or evaluation.

Before you choose

• Pricing and commercial terms are not published; confirm them with Labelbox.
• Environment quality depends on task design, verification, and difficulty progression; poorly designed rewards can teach models shortcuts.

Listed under

• Coding
• Sandboxes and infrastructure
• AI agents
• AI training data
• Data labeling
• Enterprise
• Enterprise AI
• Evaluation

Get started

Visitors are directed to contact Labelbox to get started.

1. Define the target domain, reward goals, and pass@k targets with Labelbox.
2. Provide or co-design task specifications with domain experts for environments such as research or software engineering.
3. Review calibrated environments and reward signals before using them in post-training.

Pricing, trial terms, and billing model are arranged directly with Labelbox.

We couldn’t find published prices for Horizon. See their pricing page ↗

Sources for this overview

Product pages, documentation, and pricing checked 29 Oct 2026. Each source supports part of the overview; availability and prices can change.

• Horizon | Labelbox ↗labelbox.com

Company website

labelbox.com: product, pricing and policy pages

YC directory mirror

Batch and company name

Social accounts

Accounts the company links, and its posts on X

Human review

Checked by a person

knowledge_base

Last checked 29 Oct 2026. Every field keeps its source and date; unknown means we couldn’t confirm it.