Today we're launching Recursion Managed Agents: AI agents for every team, working inside the apps your business already runs.
Give an agent a goal, not a script. It starts on a schedule, an alert, or a message, brings in specialists as needed, and hands back finished work with the evidence and a grade against your bar. Every graded run feeds its memory, so the next one starts smarter, with no retraining.
We built it to run Labelbox first. More than 50 agents now work across the company, from engineering and research to finance, sales, and procurement. We pick the right model for each job and switch when a better or cheaper one ships. What the agents learn stays with us. Our token spend has fallen from about $2 million to $1.2 million a month, while we've gotten significantly more work done.
Start close to the work you do with Labelbox:
- A labeling batch is done: export it, train your next model version on a GPU, and compare it to the last.
- A new checkpoint lands: rerun every eval and block regressions.
- A new dataset arrives: check it for duplicates, leakage, and label drift before training.
Then go further, across engineering and research:
- Engineering: build features from tickets, fix bugs with regression tests, upgrade dependencies, retire deprecated APIs.
- Reliability and security: investigate incidents as alerts fire, write postmortems, security-review every pull request, patch new vulnerabilities everywhere.
- Research: run sweeps and ablations, benchmark models overnight, turn production failures into evals, reproduce key results from new papers.
If it's knowledge work your team repeats, an agent can take it on for days or weeks, across thousands of sessions at once. When something fails, it can pick up where it left off. Each session runs in an isolated sandbox with only the tools you grant.
Pick one job your team does every week. Give an agent the goal, the tools, and what good looks like. For a limited time, new teams get $20 in free credits to get started.
Read the full post to learn how Recursion Managed Agents turn recurring work into agents that improve with every run.
