Your attention needs a chief of staff, not another agent

What it does

Chief is a local-first orchestration layer that ingests events from agents, CI pipelines, RSS feeds, and heartbeats. It runs a three-stage pipeline—hard rules, similarity matching, and an LLM judge—to route each event into one of four buckets: interrupt you, batch into a digest, dispatch to another agent with verification, or curate into memory. Everything lives in a SQLite file and markdown under ~/.chief; there is no cloud dependency or telemetry.

The interesting bit

The project treats trust as a measurable engineering problem rather than a vibe. It ships with 341 offline tests, a 200-case golden dataset, and a 100-user cohort benchmark that proves exactly when its learning loop fails—specifically, that 36% of simulated users hit a mathematically provable ceiling where feedback cannot lift a quiet topic over a scene’s threshold. It also runs a seven-day “shadow mode” where it never actually interrupts, only showing what it would have done until you grade enough calls to graduate it.

Key highlights

• Three-stage funnel: µs-level hard rules kill the noisiest 25% before they ever reach an LLM; only unfamiliar survivors hit the judge.
• Scene-aware timing: The same event can ring during deep work or wait in a 2 AM digest depending on your context.
• Verified dispatch: When Chief hands work to an agent, it verifies the result with an acceptance command or a second LLM opinion before reporting “done.”
• Explainable by construction: Every decision carries its reason, five scored components, matched rules, and per-decision USD cost; chief trace replays the full chain.
• Graceful degradation: If the LLM backend dies, Chief falls back to rules-only conservative routing and auto-heals when the backend returns—it never interrupts blind.

Caveats

• The learning benchmark shows only 64% of simulated users converge; the README is admirably upfront that the rest are provably capped by the weighting formula, but real-world convergence may vary.
• It requires Python 3.12+, which may narrow adoption.
• The “projected judgment cost” of $0.104 per 1,000 events is pegged to DeepSeek list prices and cache-aware assumptions; your mileage will vary with other backends.

Verdict

Chief is for developers already drowning in agent heartbeats, CI noise, and RSS alerts who want an auditable, local-first filter with falsifiable claims. If your notification load is light or you prefer cloud-hosted black-box triage, this is overkill.