Parallel AI work is organized in different workspaces
GitHub Copilot adds tools for organizing multiple concurrent sessions, while Instacart's Blueberry prepares context and a root-cause hypothesis with about 10 subagents after an incident alert.
Artificial Intelligence··Morning
Two forms of parallel work
Two accepted news cards describe tools that organize parallel work in different workspaces. In its Copilot notes for the week of August 3, GitHub highlights changes for running more than one session at a time. The Copilot app adds direct access to shared sessions, while the `/side` command lets a user ask a side question without interrupting the main job. The command-line tool also adds a sessions sidebar with keyboard shortcuts for managing multiple concurrent sessions. By contrast, Instacart's Blueberry runs about 10 subagents in parallel after an alert for a production incident. The shared element is the organization of work that does not fit into one linear sequence. Still, the cards do not describe the same product or a shared technical arrangement: GitHub's card concerns the flow of development sessions, while Instacart's concerns the first minutes of an incident investigation.[1], [2]
Blueberry prepares the opening incident investigation
On the Blueberry side, parallel work is used in the incident investigation that begins after an alert. Instacart says the system runs about 10 subagents at once and posts a grounded root-cause hypothesis in the engineers' Slack thread, typically within about three minutes. The system gathers information, generates hypotheses, and supports debugging. Instacart reports roughly 25,000 diagnostic passes in April across more than 270 Slack channels. The company says that grounding the system in more than 14 years of incident history raised diagnostic accuracy from the mid-60 per cent range to the high 90 per cent range. These measures are company-reported, and no independent measurement has been published. Blueberry makes no automatic production changes; diagnosis, mitigation, and remediation decisions remain with the engineer. The concurrent sessions in GitHub's card organize development work across sessions, while Blueberry's subagents prepare the opening output of an incident investigation.[2], [1]
Decision points and scope remain distinct
The two uses also have different relationships to human decisions. GitHub's card describes a side chat that preserves the primary conversation's context, access to sessions, and workspaces for separate conversations. Instacart's card describes an incident-investigation system that gathers information, generates hypotheses, and supports debugging; diagnosis, mitigation, and remediation decisions remain with the engineer. Blueberry makes no automatic production changes. Instacart reports roughly 25,000 diagnostic passes in April across more than 270 Slack channels. The company also says that grounding in more than 14 years of incident history raised diagnostic accuracy from the mid-60 per cent range to the high 90 per cent range. Those are company-reported measures, and no independent measurement has been published. The comparison does not establish either system's performance or show that one is better than the other. What they share is the visible organization of parallel work; what differs is whether that organization supports development sessions or the opening investigation of an incident, with Blueberry's decisions remaining with the engineer.[1], [2]
