LangWatch
Sign Up
← Changelog

A Governed SQL Workbench, and the Go SDK

Write SQL against your own data in a governed workbench and save the charts you build. The Go SDK ships with native instrumentations and a REST client.

LangWatch Team · August 16, 2026 · 3.13.0

What You Can Do Now

Query your data and keep the charts

  • A governed SQL workbench, full stack: write queries against your data, save workbench charts, and reopen them later.

  • Saved charts are reachable over the REST surface with a project API key, and the time-window contract is owned by the surface.

Instrument Go

  • The Go SDK ships native instrumentations, a REST client, and gen_ai-first telemetry.

Follow your coding agents

  • A Sessions screen with full-session terminal replay, and conversations restored to their proper names.

  • The pull requests page reaches GitHub parity: sorting, search, a date filter and numbered pagination.

  • Pull requests link the moment GitHub announces them, history is kept, and GitHub Enterprise Server is supported.

  • Coding-agent instrumentation is project-scoped, with lazy virtual keys.

Route and bill gateway traffic

  • Model tiers, a routing policy editor, and an allowlist that now covers aliases.

  • Per-key budget bars, one-line budget scopes, and a routing null glyph.

  • Audio and realtime usage is now counted for billing. Before this, audio calls billed as zero.

  • View the traces behind a virtual key, and manage teams and projects from both SDKs.

  • Webhooks can deliver to SQS as a per-endpoint destination.

Provision and govern

  • A full management API surface with CLI, a SCIM reference, and self-hosted provisioning.

  • A unified authorization engine: one registry, one resolver, every principal.

  • Pull Databricks Genie activity as who-asked-what visibility records, and create the Anthropic Admin source from the UI.

Automations and scenarios

  • A merged automation authoring wizard and one unified list, plus retention, overflow aggregation and plan-tiered action ceilings.

  • Scenario run parameters with HTTP secret references, and a parameters editor that opens with a row ready to fill.

  • Judge n-way target comparisons from the experiment SDKs.

  • Scenario judges gained minTurns, a guaranteed-turns floor the judge cannot end early, and an inconclusive finish now continues instead of failing the run.

Verdicts that match what actually ran

  • Evaluation display and export no longer show a verdict for an evaluation that never ran.

  • A crash in the scenario runner no longer shows up as the run's verdict, and a row the judge could not decide says why.

  • A deleted span no longer reaches a dataset, and removals are shown in the editor.

Storage and traces

  • Azure storage parity with safe provider migration, and Azure Blob sign-in without a shared account key.

  • AWS Bedrock Converse payloads are extracted into trace input and output.

Screenshots

The coding-agent Sessions screen.

screenshot

Pull requests, at GitHub parity.

screenshot

Gateway model tiers and routing.

screenshot

Audio and realtime quantities reaching billing reconciliation.

screenshot

Scenario run parameters.

screenshot

N-way target comparison results from the experiment SDKs.

screenshot

Notes

  • Shipped across langwatch 3.12.0 and 3.13.0, plus agent-plugin 1.0.0, typescript-sdk 1.5.0, and Scenario 1.1.1 for both JavaScript and Python.

  • The governed SQL workbench has no screenshot in this draft; worth adding one before publishing.

  • Under the hood: continuous profiling with Pyroscope, simulations moved onto the process-manager substrate, Prisma 7 and tRPC 11, and a batch of ClickHouse audit fixes including decimal money sums.