Start Here
Curated reading paths through everything we’ve published. Each article is a field note from running AI agent teams in production - real costs, real failures, real infrastructure decisions.
Where should I start?
If you’re a solo founder running agents and want to know what it actually costs and how to keep them from wasting money, start with what it costs and kill discipline.
If you’re a team adopting AI coding tools and want protocols that prevent chaos, start with multi-agent team protocols and sessions.
If you’re just exploring, read The System first for the operating model, then come back here for the field notes.
What It Costs
The first question everyone asks. Here is the full breakdown - every line item, no rounding, no “contact us for pricing.”
-
What Running Multiple Ventures with AI Agents Actually Costs - The complete monthly bill: infrastructure, AI subscriptions, hardware amortization, domains. About $450/month for several active projects.
-
Why We Built a Development Lab Instead of a Product - Why shared infrastructure across ventures changes the math for a solo founder, and when it makes sense to build the shop before the product.
Keeping Agents Reliable
AI agents are optimistic by default. They will churn on unsolvable problems for hours without escalating. These articles cover the operational patterns that prevent that.
-
Kill Discipline for AI Agent Teams - Five mandatory stop rules that force agents to escalate instead of spiral. The most important operational pattern we’ve found.
-
Multi-Agent Team Protocols Without Chaos - How to coordinate dev agents, PM agents, an advisor, and a human captain without them stepping on each other. Namespaced labels, QA grading, explicit role boundaries.
-
Multi-Model Code Review - Why One AI Isn’t Enough - What happens when you send the same code through multiple AI models with different strengths. Each model catches things the others miss.
-
Sessions as First-Class Citizens - How to give agent sessions the same reliability guarantees as distributed systems: heartbeats, idempotent handoffs, crash recovery.
-
From Monolith to Microworker - Kill discipline applied to infrastructure. We deleted a 3,234-line worker, its database, and its storage bucket. Nothing noticed it was gone.
Context Management
The hardest problem in multi-agent development: every session starts cold. These articles cover the system that gives agents memory across sessions and machines.
-
How We Built an Agent Context Management System - How every agent session gets immediate access to handoff state, enterprise knowledge, the work queue, and operational docs - on any machine.
-
96% Token Reduction - Lazy-Loading Agent Context - Session startup was consuming 45,000-71,000 tokens before useful work. How we cut it to 3,000 by switching to an index-and-fetch pattern.
-
Building an MCP Server for Workflow Orchestration - How to bridge AI coding CLIs to a custom backend API with typed, validated tools. The transition from fragile bash scripts to a real protocol.
-
Documentation as Operational Infrastructure - Why stale docs are actively dangerous when your developers follow instructions literally, and how to make documentation self-healing.
Infrastructure
The physical and logical infrastructure that makes multi-agent development work at the solo-founder scale.
-
Fleet Management for One Person - How to manage a distributed dev fleet with Tailscale, idempotent bootstrap scripts, and SSH mesh networking. Adding a machine takes minutes.
-
Secrets Injection at Agent Launch Time - One command, right secrets, no files on disk. A CLI launcher that scans the repo and injects credentials from Infisical at launch time.
-
One Monorepo, Multiple Ventures - How a JSON venture registry with capability flags lets a single monorepo serve multiple products with shared tooling and enforced boundaries.
-
Staging Environments for AI Agents - When your “developers” execute
wrangler deployliterally, you need a gate between development and production. -
Building a Dark-Theme Design System with Tailwind v4 - The technical decisions behind the site you’re reading now. CSS custom properties, Tailwind v4 theme config, dark-first design.
Ship Log and RSS
Short-form field notes published after each working session. Unpolished, timestamped, occasionally surprising. Browse the full ship log.
New articles publish roughly weekly. The ship log updates more frequently. Subscribe via RSS feed or articles-only RSS.
These articles document how we work. For the principles behind it, read The System. For the problems we haven’t solved yet, see Open Problems.