Blog

Writing

Thoughts on AI systems design, agentic orchestration patterns, and engineering practice.

AI-native venture leverage: 1 operator × 3 companies × $200/mo Claude Max

What it actually looks like to run three businesses simultaneously as a single operator using AI as the execution layer. Three companies, one operator, 18 months of compounding agentic practice.

Read essay →

Wiki-as-substrate: typed authority boundaries that scale across brands

Most wikis are documentation graveyards. This one is a live operational substrate — with typed write authority, daily distillation, and structural validation. Here's what makes the difference.

Read essay →

LLM Council critique loops: 3-vendor fan-out + chairman synthesis as a strategic substrate

Why a single model's strategic advice is a single point of failure — and how routing business questions to GPT-5, Grok, and Gemini before synthesizing with Claude produces better decisions.

Read essay →

9-repo direct-to-main: how /consolidate keeps the network in sync

No PRs. No long-lived branches on origin. Every repo pushes direct-to-main. Here's why that works for a single-operator multi-agent system — and what /consolidate actually does.

Read essay →

Verifier agent + auto-bug-filing: removing operator friction from the loop

The pattern that let me stop reviewing every merged ticket manually: a verifier subagent that checks acceptance criteria, reports verdict, and auto-files any anomalies it finds — whether or not they're in scope.

Read essay →

Slot-pool dispatcher: replacing wave-barrier latency with 1-in-1-out parallel

Why dispatching N tickets as a wave and waiting for the slowest one to finish is the wrong mental model — and what 1-in-1-out slot management actually looks like in production.

Read essay →

Spec-driven execution: tickets with files_predicted + acceptance_criteria as the FSM gate

The pattern that turned a vague ticket queue into a reliable dispatch system: two required fields that structurally gate quality instead of hoping the executor infers your intent.

Read essay →

The 18-day Claude Max burst: what compounding agentic-SWE practice unlocks

1,005 tickets merged. 1,458 processed. 52 agents deployed across 9 repos. What actually happens when you run a production multi-agent OS at full throttle for 18 consecutive days.

Read essay →

Why I built my own agent framework before Claude Max

September 2024. No production agent frameworks existed. GPT-4o was the frontier model. Here's what that forced me to learn — and why it mattered 18 months later.

Read essay →

Agentic Orchestration Design Patterns

Five patterns I learned building a multi-agent AI operating system that runs 7 businesses in parallel: hooks as trust boundaries, agent contracts as authority delegation, hybrid orchestration, distillation as compounding memory, and intelligent cadence.

Read essay →