Skip to main content

​
Your first investigation (~5 min)

Prerequisite: at least one incident in your account, and AI SRE activated in the console (subscription requirements and billing are in Activation and billing). Without an incident, start with “Growing it into your SRE” below.
1

Start a new session

Go to AI SRE and click New chat in the left sidebar.
2

Bring in the incident context

Type @ in the composer to open incident search, then pick the incident you want to investigate — it’s inserted as a reference chip. You can also click the Investigate an incident chip below the composer. It fills the input with a ready-made prompt, and you can append an @ incident reference before sending.
3

Send your instruction

Send something like:
4

Watch the investigation unfold

The agent streams its plan, tool calls, and intermediate findings, then delivers a conclusion backed by evidence.
You’re done when: you’ve received a reply with an investigation plan, a trail of tool calls, and an evidence-backed conclusion — and you can keep asking follow-up questions about it. If it doesn’t work, check:
  • @ doesn’t find the incident — confirm it actually exists in this account and the name or ID is spelled correctly;
  • No response for a while, or it’s stuck on “environment initialization” — the cloud Sandbox or Runner is probably still starting up; give it a moment. For persistent unresponsiveness, see the troubleshooting section in Environments;
  • The reply only has a conclusion, no plan: ask “give me an investigation plan first, then evidence for each point” and the agent will follow up;
  • You can’t find the AI SRE entry at all — the entry follows role permissions; ask your account admin to confirm your role has AI SRE access.
Next: for streaming, cancellation, forking a session, and context compaction, see Chat.

​
Other entry points

Besides Chat, there are two more:

​
Automations

For work that shouldn’t need a human watching it — scheduled inspections, weekly insight reports, alert-noise cleanup. Go to AI SRE → Automations and create one from a built-in template (On-call Incident Analysis, Alert Governance, Weekly On-call Review, and more) or build one from scratch. Each run happens in a hidden session in the background, and the result lands in the run history where you can open the full trail any time. → See Automations

​
IM (@ mentions + automatic war room diagnosis)

Where incidents actually happen. In any Slack, Feishu, DingTalk, or WeCom group with the bot connected, @ AI SRE to start or continue an investigation — it replies in-thread. Open a war room for an incident and, when an automation rule with the War room created trigger matches it, AI SRE automatically runs an initial diagnosis and posts the conclusion back. → See IM

​
Growing it into your SRE (as needed)

Your first investigation is the starting point. From here, you can make analysis automatic and keep sharpening its diagnoses. Diagnosis quality depends on how much AI SRE knows about your systems and how much real data it can reach. All of this builds up over time, not all at once.

​
Get every new incident analyzed automatically

Once manual investigations feel routine, make them the default. Go to AI SRE → Automations and create a rule from the built-in On-call Incident Analysis template. Enable the On-call incident trigger card, and pick the channels and severities to watch. From then on, every matching new incident gets an automatic investigation. AI SRE writes one summary comment back to the incident, with the conclusion first and the full session link at the end. The incident’s existing notification channels deliver it, so responders see the first analysis in IM. Configuration details: Automations · On-call Incident Trigger. This is independent of war room auto-diagnosis: this path fires for every matching new incident and writes back to the incident itself. War room diagnosis is driven by an automation rule with the War room created trigger — it runs when you open a war room for an incident and the rule matches, and posts into the war room group (see IM).

​
Build team knowledge

If the agent has to guess “what service is this,” “who owns it,” and “how was this handled last time” on every investigation, it won’t get much sharper. Type /init in a session, and the agent interviews you to map out your services, runbooks, on-call routing, and more. It confirms each write before saving, and future sessions load the result automatically. See the full interview flow in Setup (/init) and the knowledge structure and best practices in Knowledge.

​
Connect real data sources

By default the agent only sees Flashduty’s own data. To let it query your actual systems — logs, metrics, code repositories, internal databases — connect an MCP server or deploy a self-hosted Runner. Publicly reachable services (observability platforms, GitHub, and similar) install and authorize in one click from the MCP directory — see MCP. VPCs, internal databases, or local commands need a self-hosted Runner deployed on a machine that can reach them; see Environments.

​
What’s next

Product Overview

What AI SRE is, its full capability set, and console navigation.

Chat

Creating and managing sessions, streaming, forking, and context compaction.

IM Platform

@ the agent in IM to troubleshoot, and learn about automatic war room diagnostics.

Environments

Cloud Sandbox vs. self-hosted Runner, deployment, and permission configuration.