Bento

Teams building AI agents can ship them easily, but trusting those agents in production is harder. Prompt changes, model swaps, and new tools often introduce silent regressions that are caught by customers rather than internal dashboards. When a regression is found, the relevant context is scattered across separate tools: traces in one, evals in another, fixes in a third, and none of it feeds back into the agent. As a result, every release starts from zero, learnings do not compound, and the value of ongoing AI investment is diminished.

Bento is a closed-loop platform for shipping self-improving agentic systems. It captures OpenTelemetry-native traces from every framework and lets users describe any failure mode in plain English, which Bento turns into a regression signal trained on the team's own traces, fired in real time, and backfilled across historical data. Alerts can also be written in plain English, with Bento grouping them into incidents and judging each as real or benign to reduce noise, and it can detect behavioral drift by pinpointing exactly what shifted in an agent's behavior. The platform also includes Artifacts (reusable, trigger-based fixes like skills, subagents, and tools that graduate from candidate to promoted status based on production performance), a living memory called the Book that records failure patterns, fixes, and outcomes in plain language, Evaluations that score every release against production history offline, in CI, and on live traffic, and Versions that make every prompt, skill, and model change versioned, diffable, and reversible so regressions can be traced to their cause.