I ran my life on a fleet of AI agents for 6 months

Practitioner 8 min read

In January 2026, I decided to stop using AI assistants and start building AI agents.

The difference sounds semantic. It is not.

An assistant answers questions. An agent takes ownership of a domain. An assistant forgets everything between sessions. An agent reads state files that encode six months of accumulated context. An assistant starts cold every morning. An agent starts with a briefing.

By July 2026, I had 133 registered projects, 28 agent packages, and 47+ cron jobs running 24/7. A CEO orchestrator routes tasks to specialized agents across 8 life domains. Agents for health, family, work, school, growth, AI projects, consulting, and life operations — each with its own state file, memory, and escalation rules.

Most of what I built is not replicable by reading a tutorial. But the operating system underneath it — the methodology, the conventions, the 4 files — that is what this essay is about.

The thing everyone building with AI is missing

Every commercial AI assistant — Lindy, Claude.ai, ChatGPT, whatever you use — has the same problem.

They do not know you.

Not really. They know your name if you told them. They might remember a preference or two. But they do not know that you have a hard rule about not committing money over $500 without sleeping on it. They do not know that you write in short punchy sentences and you will cringe at “I hope this email finds you well.” They do not know that Sundays are sacred and no work messages should go out.

Every session, you explain yourself again. Every time an agent makes a recommendation, it is making it in a vacuum.

The commercial players are building toward this. They call it “memory.” But their memory lives in their system, in their format, behind their paywall. Switch providers — start over.

What I built instead: a personal context system that lives in markdown files on my machine. Four files. Two hours to set up. Zero vendor lock-in.

The 4 PCA Files

I call them Personal Context Artifacts. Four files that answer four questions:

File 1: pca-wiki.md — Who am I?

The canonical “who is this person” reference. Background, family, life domains, current projects, the 20 people most relevant to your work right now, your current life-state. Any agent that drafts as you or makes decisions on your behalf reads this first.

The discipline: update it on major life events. It is not a journal — it is a reference. Think Wikipedia page for yourself.

File 2: pca-mental-models.md — How do I decide?

Your decision-making priors. The beliefs and frames you apply when you do not have time to think.

Things like: “I prioritize family over work when the cost is reversible. If the cost is irreversible, family wins outright.” “I prioritize shipping at the small scale, perfection at the high-stakes scale.” “I prefer 80%-now over 100%-later for anything reversible.”

When an agent makes a recommendation, it reasons from these priors. When the recommendation is wrong, it is usually because a prior is wrong or missing. Fix the prior, not the agent.

File 3: pca-voice.md — How do I write?

The thing that makes AI-written text feel wrong is that it does not sound like you. Voice solves this.

The format that worked: 10 examples of text I would actually send, 10 examples of text I would never send, and channel-specific rules (email vs. Slack vs. formal document). Real examples beat abstract instructions.

After building this file, AI drafts consistently sound like me on the first try. Before this file, I rewrote every draft.

File 4: pca-protocols.md — What are my hard rules?

Non-negotiables the agents must never violate. Time-rules (no calls before 9am, no deliverables on Sundays). Relationship rules (never commit anyone else’s time without asking, never share someone’s private information). Financial rules. Communication rules.

These are P0 constraints. Violations are not optimizations to make — they are errors to flag.

The 5 Skills

Skills are slash commands Claude Code runs on request. They are the operational layer on top of the four files.

Open Loops — the skill that changed how I work more than any other. Before any meeting, any conversation, any email thread goes stale, the skill extracts every commitment, every deliverable, every “I will get back to you.” It creates a dated, named open loop entry. The loops do not close until someone marks them closed.

Two weeks into using this, I stopped dropping balls. That had never happened before.

Watchers — the primitive between a one-time todo and a standing cron job. “Tell me when the contractor replies.” “Notify me when this file appears.” “Remind me 3 days before this invoice is due.” Conditions that fire once when true, then stop. A watcher is a durable promise the machine keeps so you do not have to.

CRM Everything — every person who touches your work gets a file. Not a Salesforce entry — a local markdown file. Created automatically from conversation context when the skill is invoked. The relationship history accumulates over time, readable by any agent, owned by you.

Morning Briefing — reads your state files, open loops, and watchers and generates a structured briefing. Priorities for today. Open loops needing action. Domain status. What would a smart chief of staff tell you in the first 10 minutes of the day?

Session Digest — at the end of a session, generates a structured summary of decisions made, work shipped, unfinished items, and context for the next session. The next session reads it first. No more starting cold.

What I learned building this

The context is the product, not the agent. The CEO orchestrator was not the impressive thing. The impressive thing was that when it reads mental-models.md, it makes recommendations I actually agree with. Architecture is table stakes. Personal context is what creates the leverage.

Formats that do not age are worth 10x templated systems. Markdown files in a git repo outlast any SaaS. The four PCA files work with any AI model, any session, any future tool. No migration ever.

The discipline of maintaining them is the hard part. The files are useless if they are stale. The habit I built: any time an agent makes a recommendation I disagree with, I ask why. Usually something in the mental models file is wrong or missing. Update the file, not the agent.

Open loops are the biggest leverage point. More than any other piece of the system, tracking commitments changed how much I got done. Extracting every “I will” and “let me get back to you” — and tracking whether they resolved — is the single biggest operational improvement I made in 2026.

The honest disclaimer

This system took six months and 133 projects to build to its current state. What I am sharing here is the distillation — the 20% that creates 80% of the value.

The four template files and five skills take an afternoon to set up. The returns start immediately and compound over months as the files get richer and the agents get better at predicting what you actually want.

The context is not a product you buy. It is a system you build. The templates are the starting point. Your receipts — your actual examples, your real decisions, your actual hard rules — are what make it yours.