What is agent-chief?

Chief is a tool that sits between a person and everything competing for their attention, such as AI agents, monitoring alerts, CI failures, and RSS feeds. Instead of every one of those things pinging the user directly, they all flow into Chief first, and Chief decides what to do with each one: interrupt the person right away, hand the task off to an agent and check that the agent actually finished it, or quietly file it away in memory for later. The decision process runs in three stages, cheapest first. Simple hard coded rules catch obvious cases in microseconds, a similarity check catches near duplicates next, and only genuinely uncertain cases get sent to a language model for a judgment call. Chief also adjusts how easily it interrupts based on what it thinks the person is doing, so the same event might ring immediately at a desk but wait quietly in a digest at two in the morning. For its first week running, it stays in a shadow mode where it only shows what it would have done, without actually interrupting anyone. Every decision Chief makes is explainable: it records the reason, several scored factors, which rules fired, and what the language model call cost, and a command lets someone replay that reasoning after the fact. Chief also learns from simple thumbs up or down feedback on its calls, adjusting its per topic behavior over time, and the project includes benchmarks showing that learning process actually improving its accuracy across many simulated users. It runs entirely on the user's own machine, storing its state in a single SQLite database file and plain markdown files, with no cloud service or telemetry involved. Getting started takes one command that runs a fully offline demo with no setup, and a persistent version can be installed afterward to connect real agents and event sources. The project is written in Python and released under the MIT license.