Braintrust

Braintrust introduced Nitro, an asynchronous query execution engine for Brainstore that separates storage reads from compute processing, cutting the number of object-storage requests and letting concurrent reads overlap without oversubscribing CPU and memory; full-text searches now run more than twice as fast on average across benchmarked customers. Nitro is automatically enabled for all SaaS and BYOC customers, and self-hosted customers must upgrade to dataplane 2.15 or later to get it.

Matters for: Teams running self-hosted Braintrust who rely on coding agents like Codex or Claude Code to run high-volume, concurrent searches over production trace history.

Braintrust now automatically surfaces recurring failure modes and opportunities across production traces via Patterns, explains likely causes of individual complex run failures via Debugger, and lets teams investigate open-ended questions on a schedule or turn findings into datasets, evaluators, and monitors through an enhanced Loop experience. Previously teams had to manually inspect traces one by one, since production agents generate more traces than developers can review efficiently.

Matters for: Teams running LLM agents in production who need to triage large volumes of traces for cost spikes, repeated tool-call failures, or unexpected successful workflows before they escalate.