Observability and SRE agent

An SRE for your voice agents, on call for every incident.

Xebit Observability watches every layer of a live voice agent: the audio, the transcription, the model, the voice, orchestration, campaigns and the infrastructure underneath. When something breaks, its SRE agent opens the incident, finds the root cause and proposes the fix. An engineer approves it before anything changes.

INC-0142 · Sev 2 · Open
Example

Slow replies on the EMI reminder agent

  1. Detected02:14

    Response time per turn above your threshold on the EMI reminder campaign. Callers are hanging up mid sentence.

  2. Correlated02:15

    Speech to text timeouts rising in the same window. Orchestration, scheduler and audio storage healthy.

  3. Root cause02:17

    Transcription provider slow in one region. Retries are stacking up in the transcription queue.

  4. Proposed fix02:18

    Route this agent to its secondary transcription provider and hold campaign retries until latency recovers.

Waiting for an engineer to approve

Approve fix Reject
Posted to the Xebit console, your Slack or Teams channel, and the on call email list.
Why voice is different

Why does a voice agent need its own observability?

Because the failure happens in the conversation, and the caller does not file a ticket. They hang up.
01

The server is green. The caller hears silence.

CPU, memory and uptime can all look healthy while a speech provider times out and every caller waits for an answer that never comes. Generic monitoring watches machines. A voice agent fails in the conversation.

02

One sentence crosses six systems.

Every turn passes through audio, transcription, the model, tool calls, the voice and the orchestration that ties them together. When a call goes wrong, the cause is usually in one layer and the symptom in another.

03

A bad ten minutes reaches every caller in them.

Campaigns dial at volume. A regression nobody notices until the morning has already reached a whole list of customers, and the follow ups, promises to pay and bookings from those calls are gone.

What it watches

Every layer a call passes through, in one place.

Signals from the call, the providers, Xebit's own modules and the infrastructure, tied to the agent and the campaign they belong to. Per call and per node metrics, transcripts, event logs and replay sit underneath all of it.

Call quality and latency

  • Response time on every turn
  • Silence and dead air
  • Dropped and abandoned calls
  • Barge in and talk over

Model and speech providers

  • LLM errors, timeouts and slowdowns
  • Speech to text, per provider
  • Text to speech, per provider
  • Which agents each provider affects

Audio and transcription

  • Audio files recorded and stored
  • Transcription jobs completing
  • Recordings and transcripts ready for replay

Orchestration, scheduler and campaigns

  • Turn handling and business rules in the call flow
  • Scheduled calls going out on time
  • Campaign progress, retries and pacing

Human in the loop

  • Handoffs reaching a person
  • Live transfers completing
  • Calls waiting on a human

Infrastructure and agent behaviour

  • Nodes, GPUs and queues
  • Failed tool calls
  • Prompt regressions after a change
  • Off script answers
How an incident runs

What happens when something goes wrong?

The SRE agent does the legwork an engineer would do at two in the morning: notice, gather, correlate, explain and recommend. The decision stays with your team.
  1. Step 01

    Detect

    Signals from every layer are checked against the thresholds you set, per agent and per campaign, while calls are running.

  2. Step 02

    Open the incident

    An incident opens with a severity, the agents and campaigns affected, when it started and a timeline that fills in as it develops.

  3. Step 03

    Correlate

    The SRE agent looks across layers in the same window: which provider, which module, which node and which campaign moved first.

  4. Step 04

    Write the RCA

    A root cause analysis in plain language, with the evidence behind it: metrics, event logs and the affected calls, ready to replay.

  5. Step 05

    Propose a fix

    A specific change or runbook step, with what it will touch and why. Not a list of things to check: one recommendation, explained.

  6. Step 06

    An engineer approves

    Nothing changes in production until a person approves it. Reject it and the incident stays open with everything the agent found.

Why approval, and not autopilotVoice agents talk to your customers, often in regulated work like collections and lending. A wrong automatic change can reach every call in a campaign before anyone looks. So the agent does the fast part, finding and explaining the problem, and a person makes the change. You get the speed without handing production to software nobody is watching.
Where it reports

Incidents reach the people who can act.

The console holds the full record. Slack, Teams and email make sure someone sees it.

Xebit console

The full incident: timeline, severity, the RCA with its evidence, the affected calls and the proposed fix to approve or reject.

Slack or Microsoft Teams

Incidents and updates posted to the channel your on call team already watches, so nobody has to be looking at a dashboard.

Email

Incident notices and the finished RCA sent to the people who need them, including the ones who never open Slack.

Every deployment

Does it work on prem and air gapped?

Yes. Observability is part of the platform, so it runs wherever the platform runs. What changes is where reports can be delivered.
Xebit Observability by deployment model
Xebit VoiceCloudOn premAir gapped
Observability runs inXebit's infrastructureYour networkYour network
Incidents and RCAs in the consoleYesYesYes
Slack, Teams and emailYesYes, where your network allowsOnly channels reachable inside your network
Who approves a fixYour engineersYour engineersYour engineers
Included with Xebit VoiceCloud and Xebit VoiceKitWhere it sits in the architecture

Observability questions

What is voice AI observability?

Voice AI observability is monitoring built around the conversation rather than the server. It follows each call through audio, transcription, the model, tool calls, the voice and orchestration, so you can see where a call went wrong, how many calls it affected and why. Xebit Observability adds an SRE agent that turns those signals into incidents and root cause analyses.

Does the SRE agent change production on its own?

No. The SRE agent detects the problem, opens the incident, correlates signals across layers, writes the root cause analysis and proposes a fix. An engineer approves the fix before anything changes. If they reject it, the incident stays open with everything the agent found, so the team can take it from there.

What does Xebit Observability monitor?

Call quality and response time per turn, silence, dropped calls and barge in. LLM, speech to text and text to speech providers. Audio files, transcription, human handoffs, orchestration, the scheduler and campaigns. Nodes, GPUs and queues. And agent behaviour: failed tool calls, prompt regressions and off script answers.

Where are incidents and root cause analyses reported?

In the Xebit console, where the full incident, the evidence and the proposed fix live, and in Slack or Microsoft Teams and by email so the on call team hears about it without watching a dashboard. On air gapped deployments, reporting goes to the channels that are reachable inside your network.

Can I use Xebit Observability with a voice stack that is not Xebit?

Not today. Observability is part of the Xebit platform and comes with Xebit VoiceCloud and Xebit VoiceKit, because it relies on signals from every module of the stack. If you run your own voice agents elsewhere, our blog covers what to monitor and how to run incidents on any stack.

SEE AN INCIDENT END TO END

Watch the SRE agent work a real incident.

We will walk you through detection, the root cause analysis and the approval step on a live agent, and show you what your team would see in Slack or Teams.