Braintrust adds application logs to its agent-tracing platform in public preview

Braintrust, which sells tooling for evaluating and observing AI agents, has released application logs in public preview. According to the company's 8 October 2026 blog post, the feature lets teams send ordinary service logs to the same platform that already holds their agent traces, and read the two together in one view.

The new part is the logs themselves. Until now Braintrust's observability centred on agent behaviour. Braintrust says logs can now come from services that run independently of any agent, such as a database or background workers, and that its Patterns tool, which looks for recurring behaviours, can be applied to them.

The company describes several capabilities. Users can search logs, pick time ranges and filter by severity, then open a record to see the metadata the application recorded. Message templates group repeated events even when values differ, so payment failures can be counted across different payment IDs while each failure's details are kept; Braintrust says its Python SDK captures templates automatically when developers use named parameters. Logs can also be queried with SQL, for example to find which errors were most frequent over the past day.

Where a log is linked to a trace, the two appear side by side. Braintrust's example is a tool request that times out: a developer can see the error and then check whether the agent retried, switched tools or returned an answer. Linking requires the operation's trace context to be included when the log is sent.

On availability, Braintrust says the preview covers all SaaS and BYOC customers, and self-hosted customers from data plane 2.14. Logs can be sent through the Python SDK (version 0.42.0 or later, per the post's setup steps) or through OpenTelemetry's OTLP protocol using an exporter or collector, with or without an active trace.

What is confirmed is limited to Braintrust's own description. This account rests on a single source, the company's launch post, and we found no independent coverage, customer testimony or benchmark of how the feature performs. Pricing, retention limits and a general-availability date are not stated in the post. Because it is a public preview, details may change.

It matters because agent failures often start in ordinary infrastructure, such as a timed-out API or a failing worker. Putting those signals next to an agent's decisions could shorten debugging for teams already using Braintrust, though that benefit is the company's claim rather than something we have verified.

Revised 16 October 2026: Guth re-read the original sources and rewrote this article. The earlier version remains available as revision 1.