Where the gain is measured

Blueberry is the AI-assisted incident response system Instacart built for its on-call engineers. When an alert is triggered it runs about 10 software agents in parallel and drops a grounded root-cause hypothesis into the Slack thread where engineers are working, within about three minutes. The company's reported figures run as follows: roughly 25,000 diagnostic passes in April across more than 270 Slack channels, a 99.9 per cent workflow success rate, more than 58,000 MCP tool dispatches and around 60 team profiles. The work being measured is the context gathering done in an incident's first minutes.[1]

Diagnostic accuracy moving from the mid-60 per cent band to the high 90 per cent band is attributed to grounding the system in more than 14 years of incident history. The limit here is plain: every figure is company-reported, no independent measurement has been published, and the labelled set against which accuracy is computed is not described. In a production setting, accuracy depends on which incidents were sampled and on who counts an answer as correct.[1]

Who still carries the pager?

This AI system makes no automatic production changes; it gathers information, generates hypotheses and supports debugging. Diagnosis, mitigation decisions and remediation remain the engineers' responsibility. Alan Wong of Instacart describes the tool as changing the on-call engineer's starting point, so that instead of an empty investigation path the engineer begins with collected context such as logs and deployments. The task that disappears is context gathering; the accountable decision stays where it was.[1]

From here on this is a distribution question, and the answer has not been published. Does faster first triage turn into fewer pages, a smaller on-call rota, shorter outages, or into more incidents handled per engineer? All four outcomes can follow from the same efficiency gain, and they produce very different working conditions. The concrete measure a reader can wait for is a published rota size and page count per engineer; the success rates the company reports do not answer that question.[1]