BentoLabs AI
Monitoring and learning layer for long-running agents
About BentoLabs AI
BentoLabs is the monitoring and learning layer for long-running agents. We detect when agents silently fail or drift from the user's goal, system prompt, or tool contracts, show affected users and root cause, and suggest the prompt, skill, or harness fix. As more teams deploy agents, keeping them reliable in production becomes mission-critical. Bento sits directly in the production loop and gives teams the operational leverage required to scale agent ecosystems without scaling human firefighting alongside them. The result is a system that turns opaque agents into agents that can be monitored, debugged, and improved continuously. The founders learned this problem at Emergent (YC S24), where they built and operated production coding agents used by 5M+ users. Abhinav was hire #1 and helped Emergent hit SWE-Bench #1 and scale from $0 to $100M ARR in just 8 months. Kaushik was hire #2, led full-stack engineering at Emergent, and was key to building the infrastructure that made production agents reliable, observable, and debuggable. Bento's self-learning engine has also lifted ARC-AGI-3 (internal) by 2.6x and Terminal-Bench 2.0 (internal) from 42.2% to 52.4% pass@1 with the same model, tools, and budget.
Public traction evidence
Each signal links to the public source used for attribution.
- LinkedIn
We’re introducing Issues by BentoLabs AI (YC P26).
We’re introducing Issues by BentoLabs AI (YC P26). Issues are recurring production problems Bento surfaces directly from your agent’s traces, logs, and runs. Instead of treating every failed run as a one-off, Bento groups similar agent failures into clear, trackable Issues into...
- LinkedIn
BentoLabs AI (YC P26) is now live on the YC Directory. YC has felt like stepping into a pressure chamber where days blur into nights, time zones stop mattering, and every week stretches you harder than the last. Somewhere between building nonstop and watching unicorn companies start using BentoLabs, things got very real, very fast. From scrolling through YC company pages to seeing ours live there now, that feeling hasn't quite sunk in yet. Excited, nervous, proud. All of it at once. And we’re dropping something very soon. Stay tuned. Check us out here 👇🏻 https://lnkd.in/eJpFeVNY …more
BentoLabs AI (YC P26) is now live on the YC Directory. YC has felt like stepping into a pressure chamber where days blur into nights, time zones stop mattering, and every week stretches you harder than the last. Somewhere between building nonstop and watching unicorn companies...
- LinkedIn
We are live on Y Combinator. Building the monitoring and learning layer for long-running agents. Go check us out if you haven't. | BentoLabs AI (YC P26)
We are live on Y Combinator. Building the monitoring and learning layer for long-running agents. Go check us out if you haven't.
- X
And we are live!
And we are live!
- X
Almost didn't document the day, my team made sure we did. Some days you just feel the shift. We booked a studio, brought in a team, and spent the day trying to capture what @BentoLabsAI actually is right now. Where we started, where we are, and where we're going. It's one thing Show more Kaushik and BentoLabs AI (YC P26)
Almost didn't document the day, my team made sure we did. Some days you just feel the shift. We booked a studio, brought in a team, and spent the day trying to capture what @BentoLabsAI actually is right now. Where we started, where we are, and where we're going. It's one thing...
- LinkedIn
Did you see 'Harness Engineering' on your feed too?
Did you see 'Harness Engineering' on your feed too? A term that suddenly blew up everywhere. It didn't have a name until February, now Hashimoto has posted on it and OpenAI has proven it at a million lines. So what does the work actually look like? Reading a hundred...
- LinkedIn
Your team has alerts for when agents fail.
Your team has alerts for when agents fail. What about when they slowly stop doing the right thing?
- LinkedIn
We ran our recursive learning layer on Terminal-Bench 2.0.
We ran our recursive learning layer on Terminal-Bench 2.0. Same agent. Same model. Same harness. Same budget. The result: Claude Sonnet went from 42.2% → 52.4%. A +10.2 percentage-point lift, significant at p < 0.05, with a 13:3 task-level win/loss ratio (internal). The only...