Labelbox

Labelbox positions itself as "teaching AI to do real work." On one hand, it builds the environments needed to train AI for frontier AI labs; on the other hand, it provides engineering and research teams with agents that can take on backlog work.

Purpose

Labelbox offers enterprises an agent system called Recursion. The official description emphasizes that Recursion is not another assistant in an editor; its agents can start on their own under conditions such as red builds, alerts, new checkpoints, or scheduled times, and continue working in the background until the task is complete, for example fixing every flaky test, keeping dependencies up to date, rerunning every evaluation, or comparing dozens of experiments at once.

Main Capabilities

The official list includes a large number of specific work tasks that can be handed to agents, covering multiple domains:

• Software engineering: Feature Builder (builds ready tickets into working features with tests, running in isolated sandboxes), Bug Fixer (reproduces new defects, locates the cause, and writes fixes and regression tests), Dependency Upgrader (upgrades dependencies when new versions are released, fixes breakage, and reruns tests), Flaky Test Fixer (reproduces flaky tests, finds races, and fixes them), Deprecation Cleanup (migrates every caller off deprecated APIs).
• Reliability and infrastructure: Incident Investigation (correlates alerts, logs, and recent deploys when an alert fires), Postmortem Writer (drafts timelines, causes, and follow-ups based on incident channels, alerts, and deploys), Preview QA (checks every preview deployment in a real browser, clicking through the flows involved in the pull request), Cloud Cost Watch (traces every spending spike back to the change behind it and writes a fix).
• Security engineering: Code Security Review (reviews every pull request and reproduces the issues found), Vulnerability Patcher (finds every service touched by a new critical vulnerability and patches it), Cloud Posture Review (checks every cloud account for misconfigurations and exposed data), Access Review (reviews who can access what, flagging access that is unused or should not exist), Threat Investigation (investigates every identity, email, and endpoint alert).
• Machine learning research: Experiment Sweep (runsDozens of experiment configurations and compare each run with the baseline), Model Trainer (launch fine-tuning runs, watch the curves, and stop bad runs early), Ablation Runner (ablate each change in a training run to find the changes that truly affect the metric), Data Quality Audit (check each new dataset version for duplicates, leakage, and label drift before training), Research Digest (read new papers and repositories on the topics you follow and rerun key results).
• Evals: Eval Regression Watch (rerun the eval suite on every model or prompt change), Benchmark Runner (run every candidate model on your benchmarks in parallel, overnight), Regression Triage (trace each eval drop to the data, code, or configuration and prove it), Eval Set Builder (turn this week's production failures into new eval cases, each with its own grader).
• R&D and Product: Simulation Sweep (run parameter sweeps on simulation code and compare each design with the baseline), Hardware CI Triage (trace each nightly firmware and validation failure to the commit that caused it), Prior-Art Search (search papers and patents before each design review and find overlaps), Prototype Builder (turn each approved spec into a clickable live prototype for the team), Product Spec Writer (write a first-draft spec from calls and tickets for requests entering the roadmap), Product Docs Updater (update the docs when a merged pull request changes how the product works), Product Close the Loop (when a requested feature ships, write to each customer who requested it), Feedback Digest (read this week's sales calls, reviews, and feature requests to find what customers repeatedly ask for).
• Finance: Forecast Refresh (rebuild the revenue forecast from pipeline, usage, and orders), Variance Analysis (explain each material variance between budget and actual line by line), Deal Desk Review (check each non-standard deal against pricing policy and margin targets), Revenue Recognition Review (read each signed contract and find the terms that change how revenue is recognized), Earnings Review (reading financial reports, documents, and conference call transcripts, and comparing the results with your estimates).
• Sales: Account Research (building briefs from CRM, past calls, documents, and news), Deal Review (reading every call and email in late-stage deals and comparing them with your sales process), Expansion Signals (observing usage, hiring, and news across accounts to find growth signals), RFP Response (drafting RFP and security questionnaire answers from approved content), Renewal Risk Review (reading usage, account history, and call notes before each renewal).
• Strategy: Competitive I (based on existing materials, this entry is truncated in the official text and does not fully provide a specific capability description).

How to Use

According to the official text, the way it works is: agents start on their own based on triggers (such as build failures, alerts, new checkpoints, new dataset versions, scheduled times, etc.), run in the background, and interact with the team's existing toolchain (such as GitHub, Linear, Vercel, Slack, Sentry, Jenkins, PagerDuty, Datadog, Google BigQuery, Databricks, Salesforce, Gong, DocuSign, Okta, etc.), ultimately delivering reviewable artifacts, such as pull requests that pass tests, preview links, root cause drafts, postmortem drafts, results tables, checkpoints, and evaluation reports.

Applicable Scenarios

Suitable for research labs with frontier AI training needs, as well as enterprises that want to hand off repetitive backlog work from teams such as software engineering, operations, security, machine learning research, evaluation, product, finance, sales, and strategy to background agents.

Limitations

The official materials do not provide clear information on pricing, charging models, deployment methods, platform availability, usage restrictions, or required prerequisites; this content is not currently provided.