What it does
Unfire is an independent technology and innovation studio. This prototype addresses a common retail after-sales problem: one case can require classification, privacy checking and a decision about whether a human must authorize a refund. Using an entirely fictional delivery-delay case, three independent Mozaik agents prepare assessments in parallel. A coordinator combines their outputs into one readable, draft-only report. A sensitive-data placeholder is flagged and any refund is routed to human approval. Nothing is sent to a retailer or payment system. The example is reproducible with Node.js, uses no customer data and incurs no inference cost. The inference runner is deliberately deterministic, not a generative AI model, so the focus is genuine concurrency and safe workflow design.
How the agents run concurrently
One Mozaik 4.2.0 runtime is created with defineRuntime<SharedCases>(). Three distinct createAgent participants, named classifier, privacy-reviewer and action-guardian, each subscribe to the message.sent event and initiate an independent runLoop on the same fictional case. Their local inference operations overlap: the instrumentation repeatedly measured maximumConcurrentAgents = 3, and automated tests require at least two simultaneous runs. A fourth observer collects three independent model.answer events and stores the findings in shared runtime state before reporting. The coordinator waits until all results arrive, preserves separate findings, and marks the outcome as a draft. Any refund requires human approval; no real payment or external customer action occurs. Local TypeScript checks and GitHub Actions verify the example.