Incident Investigation, Started Before On-Call Joins
Ask an agent to analyze logs and metrics from your connected tools, plus recent changes, and separate confirmed evidence from possible causes. Provide any evidence it cannot access directly.
The First Minutes Go to Gathering Context
An alert wakes someone up, and the first stretch of the incident goes to opening dashboards, scrolling logs, and checking what shipped recently. The on-call engineer has to gather that context before deciding what to do.
An Agent Investigates, a Person Decides
Start an agent from an alert event. Ask it to inspect recent changes, relevant code, and configuration, compare its findings with your runbooks, and draft an investigation summary. Limit its access to the systems it needs to inspect. Keep rollback and mitigation decisions with the on-call engineer.
checkout-api error rate
8.4%Error spike started 14:02, right after deploy #812. Rollback plan drafted for on-call.
What Teams Report
9.4x
growth in weekly requests across Uber's agent tools, February to August 2026
Uber runs managed agents that triage on-call alerts, and measures them by cost per alert, mean time to recovery, and alerts triaged.
Uber's software factory postCommon Questions
What starts an investigation?
A PostHog log-alert event can start an automation. Other monitoring systems can send an authenticated event to Felan's team webhook endpoint. You can also start an agent on a schedule or by hand.
Can the agent change production?
Do not give an investigation agent production write access. Use read-only credentials and keep production changes under human control.
What does the on-call engineer get?
Ask for an analysis that separates confirmed evidence from hypotheses, proposes next checks, and includes a draft incident update.
Which models can it use?
Any major provider, plus local and private models. Bring your own keys or subscription, and there's no markup on tokens.
Can it run on our own infrastructure?
Yes. Use Felan's managed cloud, or run Felan on your own infrastructure with the Enterprise plan.
Related: Site reliability engineering docs, Isolated by default, Always on.
Keep Your Attention for the Hard Problems
Start with one repo and one recurring task.