Introducing Confident Monitor
Confident AI
Docs
Pricing
Products
LLM Evaluation
LLM Evaluation

Benchmark LLM systems with research-backed metrics.

LLM Observability
LLM Observability

Trace, monitor, and alert on production LLM systems.

AI Red Teaming
AI Red Teaming

Stress-test LLM apps against adversarial attacks.

AI Governance
AI Governance

Enforce AI standards and controls across teams.

Contact salesSign Up
Contact salesSign Up

October 23, 2026

  • Observability
  • Integrations
  • Security
  • Quality of Life

Off the Rails

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

We went off the rails, on purpose. Confident AI now runs fully on AWS. Railway got us a long way (thanks for the ride!), but we'd outgrown the station, and the new home means a steadier platform with fewer bumps along the way. With the move behind us, we made some noise: audio classifiers now listen to your voice agents directly, the Observatory audio player marks every issue right on the timeline so you can scrub back and replay the exact moment things went sideways, and Pipecat audio flows straight into your traces. The only thing allowed to go off the rails now is your agent, and at least you'll hear it coming.

Off the Rails

Added

  • Span Classifiers - Classifiers now run on individual spans, not just whole traces and threads. Point one at the retriever, the planner, or that one subagent you have trust issues with. Finally, a classifier with a real attention span.
  • Workflow Chaining for Spans & Threads - Spans and threads can now kick off workflows, so classifying a single step or a whole conversation sets off whatever comes next. Your automations are no longer hanging by a thread.
  • Audio Classifiers - Classifiers can now listen to audio directly, catching the interruptions, awkward silences, and tone problems a transcript never shows. Sounds like a sound investment.
  • Audio Review Step for Audio Classifiers - Audio classifiers get a review step, so a human can listen to the clip and confirm the label before it sticks. Everyone deserves a fair hearing.
  • Revamped Audio Player in Observatory - The audio player has been rebuilt with issue timestamps right on the timeline and a dragger you can scrub back and forth to replay any moment. Jump straight to the exact second your agent said something regrettable, then replay it as many times as it takes. Everything scrubs up nicely.
  • Multiple Issues per Classification - Classifiers with the issue tag can now flag every issue in a trace, span, or thread instead of stopping at the first one. Your agent has 99 problems, and now you can see all of them. We'll issue a full report.
  • Evidence in Failure Modes - Failure modes now show the evidence behind them, so you can see exactly which traces and moments put them there. Exhibit A, B, and C, all in one place. The prosecution rests.
  • Scheduled Failure Modes & Code-Aware Write-Ups - Failure modes can now run on a schedule, and their write-ups know your code, so they can point at what to fix instead of just what broke. Failing on schedule, like a true professional. Every suspect, lined up by line number.
  • GitHub Integration for Classifiers - Classifiers can now connect to your GitHub repository, so they see the code behind the behavior. Now they finally git the full picture.
  • Classifiers Page, Collections & Tasks - Classifiers have a new page of their own, can be grouped into collections, and run as classifier tasks. Organized, scheduled, and very well behaved. A true class act.
  • Credits & Trace Storage (GB-Months) with Self-Serve Upgrades - Usage is now measured in credits and trace storage by the GB-month, and you can upgrade your plan yourself without booking a call. Credit where credit is due, and storage you can bank on.
  • BYE Support for SIP Voice Connections - SIP voice connections now handle the BYE end-call event, so calls end cleanly instead of lingering as zombie sessions. Your voice agent finally knows how to say goodBYE.
  • Media Encryption & SRTP for SIP Voice Connections - Audio on SIP voice connections can now be encrypted end to end over SRTP. What's said on the call stays on the call. Loose lips sink ships, but encrypted SIPs don't.
  • Pipecat Audio Tracing - Pipecat pipelines can now send their audio to traces, so you can hear exactly what happened alongside every span. Every word, every pause, purr-fectly traced.
  • Custom Token Request Auth for AI Connections - AI connections can now authenticate by fetching tokens from your own mTLS token service. Mutual TLS, mutual trust. Consider it a token of our appreciation.
  • Refreshed Dashboard - The dashboard has a new look, cleaner and easier to scan, so the numbers you care about are front and center. Same data, better view from the window seat.
  • New Onboarding Wizard - A new step-by-step wizard walks you from sign-up to your first traces and evals without the guesswork. Tickets ready, the smoothest boarding on the line. All aboard.
  • Agent Skills for Onboarding - Your coding agent can now pick up Confident AI skills and handle the setup for you, from installing the SDK to sending your first trace. Let your agent do the onboarding; it's a real skill issue in the best way.

Changed

  • Fully on AWS - The platform now runs entirely on AWS, for steadier deploys and fewer surprises. Railway was a great ride while it lasted; we just needed a longer track.

That's the drop for this week. 18 features and one big move, and we still made every connection. See you next Friday. Off to the next station!

Built byConfident AI