I Built an AI Chief of Staff

By Doneyli De Jesus, published March 2026 in "Signal Over Noise" newsletter.

The Problem

Overwhelming information management across four email accounts, multiple calendars, family activities, and professional obligations. Critical items being missed amid noise.

Core Architecture Patterns

Two-Tier Processing

• Tier 1 (30-minute intervals): Rule-based scanning with zero LLM cost for urgent detection
• Tier 2 (daily 5 PM run): LLM classification and draft generation for accumulated emails

This reduced API costs by ~80% by reserving expensive model calls for judgment-requiring tasks.

Graduated Autonomy

Trust operates across three escalating levels: - Level 1: Full human approval required - Level 2: Auto-send under strict conditions (edit distance <10%, confidence >0.9) - Level 3: Autonomous sending with hardcoded exceptions for VIP/family contacts

Trust is a rolling 90-day window, not a permanent achievement. Can be revoked based on performance degradation.

Three-Layer Memory

1. Observations: Zero-cost structural logging of corrections and preferences
2. Memories: Daily LLM synthesis creating durable institutional knowledge
3. Retrieval: Full-text search (BM25) capped at ~550 tokens per query

Memory decay prevents context pollution -- older memories fade unless repeatedly accessed.

Key Quote

"When your agent sends an email it shouldn't have at 2 AM, 'I don't know how this layer works' is not acceptable."

Technical Implementation

• Repurposed 2022 M1 MacBook Pro as dedicated agent server
• Docker for ClickHouse, Langfuse, Postgres; Python with Poetry
• macOS launchd for scheduling (7 background jobs)
• Tailscale for secure access; Signal for encrypted notifications
• Monthly LLM: ~$100 (Claude Max subscription supporting 26 agents)
• ~50 emails daily per tenant with >80% send rate accuracy

Key Stats

313 commits, 43,000 lines of Python. 25 database tables, 133 test files (52 security tests). ~2,000 active observations, ~400 synthesized memories. Tier 1 alerts average 3-5 urgent catches daily in <10 seconds.

Actionable Patterns

1. Deterministic Fast Paths: Route 80% through rule-based systems; reserve LLM for the 20% requiring judgment
2. Earned Trust: Never deploy agents with full autonomy; implement measurable graduation criteria
3. Bounded Memory: Three-layer capture-synthesize-retrieve with explicit decay functions