Triage

Triage is an AI SRE for teams with no SRE. Most small teams have no one whose job is reliability: when production breaks, whoever wrote the code gets paged, if anyone does, and starts reading logs half asleep. Triage does that first hour of work, and hands every decision that changes something to a person. It watches a service three ways: SLO burn-rate alerts and traffic rules on its OpenTelemetry data, checks on public URLs from the outside, and a regular audit of its dependencies. When something breaks, a tool-calling agent on Meta's models investigates the service's logs, traces, metrics, database statistics, deploy history, git history and source, and writes a report in which every claim cites the evidence behind it. Triage then proposes one reversible mitigation (turn a flag off, roll a release back) and applies it once a person approves. It verifies the recovery holds before resolving, and proves a permanent fix in a sandbox: a regression test that fails on the current code and passes with the patch, plus a typecheck. An approved fix becomes a draft pull request. People can ask follow-up questions (typed, spoken or with a screenshot), get a blameless postmortem, and read it all from Muse Code over MCP. In the live run shown in the demo video, a connection leak went from fault to cited root cause in about two minutes (74 s of investigation, on Muse Spark Open served on Baseten), the proven fix was ready 31 s later, and the incident resolved after a person approved each change. Prior work: built solo during the build window, Oct 26 to Nov 4, 2026. The only earlier work is boilerplate for the Shopfront demo store, started around Oct 24–25.