+10.2 pp pass@1 on Terminal-Bench 2.0 with a recursive learning layer: same agent, same model, same budget
+10.2 pp pass@1 on TB2: same Claude Sonnet 4.5, same harness, same budget — a controlled experiment with verified artifacts.

Monitoring and learning layer for long-running agents.
bentolabs.aiCrawled Oct 15, 2026
+10.2 pp pass@1 on TB2: same Claude Sonnet 4.5, same harness, same budget — a controlled experiment with verified artifacts.
Self-learning on ARC-AGI-3: same agent, same budget, 2.6× score, 34% cheaper per solve, and three first-ever solves.
Essays and deep-dives from the BentoLabs team on AI agent observability, self-learning systems, evaluation, and closed-loop production.
Harness engineering is the next evolution after prompt and context engineering — and most enterprise teams still don't have anyone whose dedicated job it is.
Why agents regress silently in production — context drift, trajectory corruption, distribution drift — and the closed-loop infrastructure that catches and fixes them durably.
AI agents that work in pilots often fail in production. Usually a diagnostic failure, not a capability one — how to find the layer that's breaking.
Terminal-Bench 2.0 explained: why its 89 production-grade tasks make it the closest public proxy to real agent reliability.
BentoLabs AI is built by engineers who scaled AI agents, saw them break in production, and decided to fix it. Backed by Y Combinator (YC P26).
BentoLabs responds to messages within a working day.
Ship AI agents that compound. Full-fidelity traces, plain-English alerts, auto-authored skills, and regression-proof evals — OpenTelemetry-native.
Bento Labs AI, Inc.'s privacy policy describes how it collects, uses, and protects information.
Bento Labs AI, Inc.'s terms of service govern use of the Bento platform.