Introducing Confident Voice: Watch the Launch Video Now

October 30, 2026

  • Evals
  • Observability
  • Human Feedback
  • Integrations
  • Quality of Life

TGIF! Thank god it's features, here's what we shipped this week:

The side nav is now nested, so every workflow finally has a home of its own. Evals, Failure Modes, Observe, Test, Annotate, Reports, and Red Teaming each live in their own section of the rail, and open into a sub-nav with only the pages that workflow needs. Observe, for example, keeps traces, signals, alerts, and dashboards together. Pick a lane and everything you need is right there. There's no place like home, and now there are seven of them. Nav-igate responsibly.

Nav a Nice Day

Added

  • Nested Side Nav - A slim rail for each workflow, and a nested sub-nav for everything inside it. Each one gets a dedicated home, so you stop scrolling past Red Teaming to reach your traces. Home is where the nav is.
  • LLM Analytics - See LLM cost per trace and your most expensive spans across all traces, with calls, average cost, and total cost for each. Flip between cost, latency, and tokens to find what's slow, what's pricey, and what's both. Now you know exactly how your budget was _span_t.
  • Dedicated Metric & Classifier Pages - Metrics and classifiers each get a page of their own, and classifiers are no longer tied to a data model. Finally out of the model's shadow and in a class of their own.
  • Validate Metrics & Classifiers on Real Data - Test a metric or classifier against actual spans, traces, or threads, and save sets of them to run validations on. No more grading on a hypothetical curve.
  • Revamped Workflows with Live Counts - Workflows have a better builder, and every node now shows live eval and classification counts. Watch the work flow in real time.
  • Journeys Public Endpoints & MCP Tools - Journeys now come with public endpoints and MCP tools, so your code and your agents can drive them too. The journey continues, now off-road.
  • Voice Metrics on Traces - Voice metrics now run on traces, so you can score how your agent sounds, not just what it says. Results you can really speak to.
  • Generate with AI for UI Transformers - Describe how a trace should look and AI writes the UI transformer for you. A real transformation, no code required.
  • Multi-Turn Goldens from Trace Ingestion - Trace dataset ingestion tasks can now insert multi-turn goldens. Every trace gets its turn.
  • Filterable Annotation Queues - Queue items are now filterable, and form response fields show up as table columns. Your reviewers can finally queue up exactly what they want.
  • Custom Model Parameters for Prompts - Prompts now support custom model parameters, so you can tune every knob your provider offers. Your prompts, your parameters, no boundaries.
  • Inference Settings & Embedding Model Config - Set inference settings for your evaluation model, and pick which embedding model the platform uses. Fine-tuned to your settings, embedded in your stack.
  • AWS SigV4 Auth & Turn-Level Metadata for AI Connections - AI connections now support AWS SigV4 auth and can parse metadata per turn. Signed, sealed, delivered, one turn at a time.
  • Agent Entrypoint Control for Failure Mode GitHub Connections - Point the failure mode agent at the right entrypoint in a monorepo or large codebase. Big repo, no problem; we'll take it from the top.
  • Per-Project Limits for Classifiers & Online Evals - Set per-project limits on classifiers and online evals so no project eats the whole budget. Know your limits, love your bill.

Changed

  • Clearer Rate Limit Errors on Metrics - When a provider rate-limits a metric, the error now says so plainly. No more cryptic messages; we're rate-ing our errors higher.
  • Worker Stability - Fixed workers crashing due to profiling. They'd been overthinking it; we stopped the profiling.
  • Zero-Downtime Rollouts & Readiness Gates - Deploys now roll out with zero downtime behind readiness gates. Ready when you are, down for nothing.
  • Leaner Queue Jobs in Redis - Queue jobs take up less room in the Redis cache. Same jobs, smaller footprint.
  • Helm Chart v1.0.0 - The Helm chart hits v1.0.0 for self-hosted deployments. Taking the helm, officially.

That's the drop for this week. Nav-igated, validated, and celebrated. See you next Friday.

Built byConfident AI