Skip to content

Demo

An alert becomes an agent session in its Slack thread, a reply continues the same session, a mention starts a new one, and the agent asks for a decision and waits for the answer. Then a team adds its agent with one Binding.

The demo, 98 seconds, narrated with captions.

The Slack shots are screen recordings of a demo environment with seeded faults and real agent output, sped up in places; a note under each shot says so. To run a sandbox like it on your own machine, follow Your first agent in Slack.

A cross-team incident

A payments release breaks checkout. Three agents, each in its own Slack channel, investigate, escalate and page, and the payments agent proposes a rollback that waits for one person to approve it.

The cross-team incident, 3 minutes, narrated with captions.

What Agent Kourier does in it:

  • One Binding per channel. checkout's agent answers in #checkout-alerts, payments' agent in #payments-ops, and an incident-commander agent in #incidents, which both teams share. See Write a Binding.
  • The alert starts the investigation. Alertmanager posts the alert to #checkout-alerts, and a bot-message trigger on that channel's Binding turns the post into an agent session in the alert's thread. See Investigate Alertmanager alerts.
  • Agents hand off through Slack. The escalation to #incidents and the page to #payments-ops are posts by a pager bot, and each of those channels' Bindings starts its agent from that bot's posts with its own trigger. See Start investigations from any bot's messages.
  • The rollback waits for a person. The payments agent's rollback tool requires approval in kagent; the payments Binding turns on tool approvals and lists the payments team's Slack user group in approverGroups, so a press from anyone outside that group is refused. See Let the agent ask questions.

The pager bot, the service catalog, the escalation tool's evidence checks and the agents' prompts belong to the demo, not to Agent Kourier. The Slack shots are screen recordings of the demo on a kind cluster, with a seeded fault and real agent output, sped up in places; a note under each shot says so. The demo's agents, MCP servers, Bindings and Slack setup are in the incident demo repository, whose README lists what it needs and how to run it.