Oncall

Correlates the alert with the deploy, the trace and the log line, proposes the rollback, and waits for a human to press it.

320 tokens/s Throughput

96.8% Accuracy

210ms Median latency

19.4M Calls per month

Reads metrics, traces and change history together; states a most-likely cause with its confidence; drafts the remediation but never executes a production change unattended.

Scope

Reads metrics, traces and change history together; states a most-likely cause with its confidence; drafts the remediation but never executes a production change unattended.

Capabilities

Alert-to-change correlation
Distributed trace reading
Rollback and fix PR drafting
Blast-radius estimation
Timeline for the post-mortem

Runs on

Nexus-Fast-8B GPT-4o Mini Claude 3.7 Sonnet

Deployed in

DevOps and cloud-native operations

Calling it from the SDK

const oncall = new NexusAgent({ role: 'sre', execute: false });

const finding = await oncall.triage({
  alert: pagerduty.incident(id),
  window: '30m',
  correlateWith: ['deploys', 'traces', 'logs']
});

Recorded exchanges

Published examples, reproduced verbatim. This page does not call a model.

Latency spike at 03:12

Asked

Checkout p99 latency went from 240ms to 4.1s eleven minutes ago. What happened?

Returned

Most likely cause, confidence moderate: the deploy of payments-service v4.18.2 at 03:07, four minutes before the spike.

The evidence: p99 rose on exactly the pods running the new revision and not on the four still on 4.18.1. Traces show the time is spent in a single database call that was previously served from cache; the diff includes a change to the cache key that would miss on every existing entry.

What argues against it: the same deploy went to staging six hours ago without this behaviour — staging has a fraction of the cache working set, so a cold-cache regression would not have shown there. That is consistent rather than contradictory.

Rollback to 4.18.1 is drafted and ready. It will take roughly 90 seconds and affects the checkout path only. I have not run it — someone should press it.

Timeline exported for the post-mortem.

SRE Incident response Observability

All agents