Skip to content

The AI on-call engineer

Production is down.Your phone rings.It’s already diagnosed.

Vigil is the AI on-call engineer that investigates first, calls you with the fix, acts only on your spoken yes, and closes the loop.

Your own private sandbox. Nothing real is harmed.

checkout · error rate

deploy 4.2.1flag off

Diagnosed

before ringing

Approval

“Do it.”

Recovery

verified live

The story

2:47 AM. Checkout is down.

Priya is on call for a food-delivery app in Bengaluru. A release went out at 2:39. At 2:43, one in five customers can no longer pay.

Today

Still broken
  1. 2:47

    CRITICAL: checkout errors 18%

    Her phone screams. Priya is fast asleep.

  2. 2:49

    Laptop, VPN, two-factor

    Still loading. Customers are still failing to pay.

  3. 2:53

    Dashboards

    Which service? Grafana, logs, traces, one tab at a time.

  4. 2:57

    Deploy history

    A payments release went out at 2:39.

  5. 2:59

    Slack

    Who shipped it? Arjun is asleep too.

  6. 3:01

    Still no fix

    About 3,000 orders have failed. Customers are posting screenshots online.

14 min

just to find the cause. The fix hasn’t even started.

With Vigil

Recovered
  1. 2:47:00

    The alert fires

    Vigil reads the metrics, recent deploys, logs and the runbook first.

  2. 2:47:20

    Her phone rings

    “Checkout is failing for 18% since 2:43, a minute after Arjun’s release. Shall I roll back?”

  3. 2:47:31

    “Yes, roll it back.”

    Blocked: the release ran an irreversible migration. Vigil explains and offers the safe fix.

  4. 2:47:44

    “Do it.”

    Feature flag new-coupon-engine switched off.

  5. 2:48:15

    Recovery verified

    Errors back to 0.4%, read from live metrics, not assumed.

  6. 2:49

    Loop closed

    Status page, #incident, follow-ups and the postmortem draft. She goes back to sleep.

~2 min

from alert to verified recovery, with a human approving every change.

Why it matters

Every minute of an outage has a price.

Most of it is spent before anyone touches the fix: waking up, logging in, finding what changed, finding who owns it.

average cost of one customer-facing incident

$794K1

average cost of one customer-facing incident

average time to resolve it

175min1

average time to resolve it

cost of one hour of downtime for over 90% of mid-size and large firms

>$300K2

cost of one hour of downtime for over 90% of mid-size and large firms

UPI outage on 12 April 2025, India’s third in 17 days

~5hours3,4

UPI outage on 12 April 2025, India’s third in 17 days

Targets slip. Teams aim for a 4.7-hour mean time to repair and actually get 7.8 hours. 59% of engineers say half their alerts aren’t useful.5

People burn out. Nearly 70% of SREs link on-call stress to burnout and to leaving their job (industry report, secondary coverage).6

How it works

It calls you with the fix, not the alert.

  1. 01

    Investigate

    Before your phone rings, Vigil reads the alert, live metrics, recent deploys and flags, logs and the runbook.

    7 read-only tools

  2. 02

    Call

    It calls with the cause and a recommended fix, not a wall of alerts. Interrupt any time; it listens.

    Agora RTC · Murf Falcon

  3. 03

    Act on your yes

    A fix runs only after a clear spoken yes, checked by the server. If it fails, Vigil says why and offers the next-safest one.

    1 gated tool

  4. 04

    Prove it, close the loop

    It verifies recovery from live metrics, then updates the status page and incident channel, files follow-ups and drafts the postmortem.

    5 follow-through tools

Safety in code

Nothing touches production without your yes.

Not a promise in a prompt. The server enforces it, logs it, and shows you every decision.

What Vigil did · INC-2481

  • 21:13:04

    rollback_deploy · payments-service

    “should we roll back?” · no spoken approval

    Blocked
  • 21:13:11

    rollback_deploy · payments-service

    “Yes, roll it back.” · irreversible migration 0087

    Failed
  • 21:13:26

    disable_feature_flag · new-coupon-engine

    “Do it.” · errors 18% → 0.4%, verified

    Done
  • One door to production

    Only one tool can change anything, and only from an allow-list for the incident at hand.

  • Your yes, checked in code

    The server matches the approval to what you actually said or typed in the last two minutes. Questions, hedges and “wait” never execute.

  • Unsafe fixes refused

    Restarting the database mid-incident is refused with the reason, and Vigil offers the runbook’s safer step instead.

  • Verified, not assumed

    Recovery is read from live metrics before Vigil tells you it worked.

  • Logs are data

    Instructions hidden inside log lines are never followed. The demo includes one, on purpose.

  • Idempotent and audited

    Repeated actions return “already done”. Every call, input, result and approval quote lands in the timeline and the postmortem.

Why voice

The fastest interface at 3 AM is a conversation.

Today’s AI SRE tools live in Slack and web dashboards. Vigil meets you where incidents actually happen: on a call.

  • Half-asleep. Hands busy.

    At 2:47 AM nobody reads twelve dashboards well. One calm voice with one clear question beats forty Slack messages.

  • Incidents already run on calls.

    “Add Arjun to the call.” Vigil pages a teammate into the same live call over Agora RTC and briefs them in two sentences.

  • Loud room? Type.

    Every call has a text box. Typed approvals are checked exactly like spoken ones.

Compare

The market is real. The voice is missing.

AI SRE is a billion-dollar category, and every leading product works in text. The voice prototypes stop at one rollback. Vigil does the whole incident.

  • Where it talks to you

    AI SRE tools
    Slack or a web UI
    Voice bots
    A phone call
    Vigil
    A live voice call over Agora RTC, with typing as a fallback
  • When it calls

    AI SRE tools
    It doesn’t
    Voice bots
    As soon as the alert fires
    Vigil
    After it has investigated, with the cause and the fix
  • Acts on production

    AI SRE tools
    Mostly read-only, or suggests
    Voice bots
    One scripted rollback
    Vigil
    Allow-listed fixes, each gated by a spoken yes checked in code
  • When the fix fails

    AI SRE tools
    A human takes over
    Voice bots
    Not covered
    Vigil
    Explains why and offers the next-safest fix
  • Proves recovery

    AI SRE tools
    Sometimes
    Voice bots
    Some re-check one dashboard
    Vigil
    Verifies from live metrics before saying it worked
  • Teammates

    AI SRE tools
    A separate bridge call
    Voice bots
    No
    Vigil
    Pages a teammate into the same live call and briefs them by voice
  • After the incident

    AI SRE tools
    Postmortem inside the tool
    Voice bots
    No
    Vigil
    Status page, incident channel, follow-up tickets, postmortem with timeline
  • Untrusted log text

    AI SRE tools
    Varies
    Voice bots
    Not handled
    Vigil
    Treated as data, with a prompt-injection test in the demo
  • AI SRE tools: Resolve, Bits AI, PagerDuty7,8,9 · Voice bots: VoiceOps, EZ OnCall, OpsPilot10,11,12

Built on

Real-time voice, real tools.

Every word you hear is Murf Falcon. Every fix is an HTTPS tool call made by Agora’s Conversational AI Engine, in the middle of the conversation.

  1. Alert

    Error rate, latency or payment success crosses its threshold.

  2. Investigate

    Read-only tools: incident, service health, logs, changes, runbooks, dependencies.

  3. Agora Conversational AI

    ARES speech recognition (en-IN) → GPT-4.1 mini → Murf Falcon 2.1 voice “Anisha”.

  4. Your phone

    Agora RTC call with live captions over RTM. Teammates join the same channel.

  • Agora Conversational AI Engine
  • Agora RTC
  • Agora RTM
  • Murf Falcon 2.1 · “Anisha” en-IN
  • GPT-4.1 mini · Agora-managed
  • ARES speech recognition · en-IN
  • 13 custom HTTPS tools
  • think API for live updates

Break production.
Safely.

Pick an incident, answer the call, and tell Vigil what to do. Your own private sandbox. Nothing real is harmed.

Open the live demo

Works in Chrome or Safari. Allow the microphone, or type instead.