SmileLikeYe/agent-chief

The core idea is blunt and correct: as agent pipelines multiply, so does the noise they generate. This project tries to sit in front of all of that — agents, alerts, monitoring feeds — and make one binary call per incoming signal: interrupt the human, or do not. The local-first architecture is the right call for this use case, because sending every alert through a cloud classifier adds latency and a dependency that defeats the point of fast interrupt routing. What makes this worth looking at rather than just bookmarking is the framing: it is not another dashboard or aggregator, it is an explicit prioritization layer. Reservation: the project is early, and the quality of the interrupt/no-interrupt decision depends entirely on how well the classification rules or model are tuned for your specific stack. Out of the box, expect to spend time calibrating it before it earns trust. -> Best for: SaaS team of 2-5 running multiple agents and drowning in signal noise