Ticket moved to ReadyLinearGitHubVercelSlack
Delivers: A tested pull request with a preview link, ready for review
Engineering
Feature Builder
Builds each ready ticket into a working feature, with tests, in its own sandbox.
Where AI learns to do real work
Labelbox builds the environments frontier labs use to train AI and the platform enterprises use to put agents to work.
Recursion· Sessions
Recursion· SessionsAgents at work in your business apps
DEAL-217 · Halden Systems · $42M acquisition · IC meets Monday
Diligence: can we sign the Halden Systems acquisition by Monday?
Read the data room, source every red flag, tie the numbers to the audit and draft the investment committee memo for the deal lead’s review.
running
7 agents · 4 models
outcome 0.93 · memo approved · 11 red flags sourced
- Box0
- Google Drive0
- Web0
- Notion0
- Salesforce0
- Slack0
coordinator · Box: Halden data room shared · 1,240 documents · IC meets Monday · plan: 4 workstreams
An acquisition’s due diligence, worked end to end.
A coordinator and five specialists read the data room in Box, check the courts and the audit, then draft the investment committee memo for review in Slack. A grader agent scores the outcome.
That was one of 144 sessions this week.
Engineering, research, finance and sales agents, started by schedules, alerts, events and API calls. Every run is graded.
Every run makes the next one better.
Graded sessions leave memories, notes and skills that the next session reads before it acts.
For enterprises
Agents for the work that keeps coming back
Recursion isn’t a chatbot. Its agents start on their own, on a schedule, an alert or an event, and work in the background until the job is done: every ready ticket built, every new bug fixed, every dependency kept current, dozens of experiments run at once. Here are a few of the jobs teams hand it.
Give it a job. Get back finished work.
Describe the job, what starts it and what a good result looks like. From then on Recursion runs it in the background: it plans the work, brings in as many specialist agents as it needs, runs them in parallel and delivers the result into your apps, with the evidence attached. Models, sandboxes, credentials, memory and grading are handled for you, so most teams use it out of the box.
You decide
Experiment Sweep
- Job
- Run every config in the spec and find the best one
- Apps
- LinearGitHubDatabricksSlack
- Starts
- Issue labeled sweep
- Rubric
- Every config runs to completion
- Each run compared with the same baseline
- The winner is rerun to confirm
Recursion runs
- ModelsAny model, chosen per agent
- SandboxesFor running code and using a browser
- CredentialsKept in a vault, scoped to each tool
- Multi-agentOne agent per task, and more for big jobs
- MemoryNotes and skills from past runs
- GradingEvery result checked against your rubric
You get
SlackSlack
48 configs ran in parallel overnight. Config 31 beats the baseline by 2.1 points, confirmed on a rerun.
- Every config runs to completion
- Each run compared with the same baseline
- The winner is rerun to confirm
Every run lands on one fleet board.
Over time
Gets better, and cheaper, the more it works.
Every graded run leaves notes and skills the next run starts from. When a job runs often enough, Recursion turns its graded runs into training environments, the kind Labelbox builds for frontier labs, trains a specialist model on them and switches over once it beats the model you use today.
5–10×lower cost per task once a specialist takes over
Specialist takes over
GradeCost per task
MemoryBetter every runSpecialist modelBetter and cheaper
For frontier AI
The data and environments frontier models learn from
Over 90% of leading US AI labs post-train and evaluate on Labelbox.
Hundreds of AI teams build with Labelbox
How Meta built GIM with Labelbox data to evaluate frontier AI reasoning
Problem
Meta needed a benchmark that remained discriminative as existing LLM evaluations saturated. The team wanted tasks grounded in practical reasoning rather than obscure knowledge or synthetic puzzles, with enough rubric detail to capture partial credit and enough quality control to support a public-private contamination diagnostic.
Solution
Labelbox produced the data foundation for GIM: 820 expert-authored problems across seven cognitive categories, including 229 multimodal items and 528 rubric-graded prompts. The work included original prompt creation, structured scoring criteria, review, expert feedback, and quality assurance, enabling Meta to calibrate a 2PL IRT model over more than 200,000 prompt-response pairs.
Result
Meta released GIM-615, calibrated item parameters, and an evaluation framework that benchmarked 22 models across 47 reporting configurations. The paper found GIM remains far from saturated, with roughly 20% of items above frontier ability, giving researchers a durable way to compare model capability, thinking budgets, and future systems.
Human preference signal for evaluating LLMs inside Vertex AI
Read story
Retail and ecommerce
Higher-quality training signal for personalized shopping AI
Read story
Tracking surgical instruments in video to advance robotic surgery
Read story
Labelbox Research
Latest work from Labelbox Research
Our applied research team publishes the benchmarks and methods we use to train and evaluate frontier models.
Built for enterprise security
Each agent gets only the access its job needs, in a workspace of its own. The model never sees your passwords or keys, and every step is recorded.
- SOC 2 Type II
- GDPR
- CCPA
- NIST 800-171
- PCI DSS
Get started with Labelbox
For enterprises
Stop prompting. Start delegating.
Your teams already ask AI for answers. Recursion takes on whole jobs: it starts on its own, works across your apps and hands back finished, graded work. Bring the job that eats your experts’ week, and we’ll set it running.
For frontier AI
Build what your models learn next
Frontier AI labs already train on Labelbox environments and expert data. Tell us where your models fall short, and we’ll scope the environments, data and evals.