what AI agents still can't do (honest)
A realistic look at what AI agents still cannot do: uncertainty, context they were never given, and judgment calls that require being you.
Cathryn Lavery · Builder & Operator · Est. 2013
For operators getting technical, and builders learning to operate.
I go deep on one hard thing at a time, then write down exactly how to do it yourself. The topic keeps changing; the method doesn’t.
My best stuff is in my email.
My goal is for it to be the most useful email you read that day.
No spam. Unsubscribe in one click.
Gone deep on: Productivity Kickstarter DeFi Business AI
Writing since
2013
13 years of annual reviews, wins, and wrong turns.
Essays in the archive
161
From Kickstarter to crypto to AI agents.
The latest
Claude Code hooks turned repeated instructions into enforced behavior. Cat explains the difference between agent context and automation guardrails.
Start here
The signature essay
How I walked away from my own company — and why I couldn't stay gone.
An employee stole $230k from my company
What you learn about people — and yourself — when the money disappears.
Why 50/50 partnerships are lazy
The math problem nobody solves before they sign.
How I built my wife a personal AI assistant
The tech took a day. Understanding how she'd actually use it took a week.
First 100 days: buying my business back from private equity
What I fixed, what they broke, and what 100 days of rebuilding taught me.
A realistic look at what AI agents still cannot do: uncertainty, context they were never given, and judgment calls that require being you.
A practical way to make a local AI agent reachable from Telegram, Slack, or another inbox while keeping the work close to your files and tools.
Cat’s accidental technical founder story: failed app handoffs, AI-assisted making, Milo, and the shift from building things to sharing them.
A flat to-do list fails when AI agents do the work. The 7-step Beads workflow I use to plan and execute projects with Claude and Codex.
Claude Code becomes an operator the moment you wire up MCPs. Six concrete non-coding workflows: SEO briefings, niche video research, PRD automation, and more.
The quiet failure mode in AI agent teams is context drift. Cat explains why stale instructions break output and how to audit them.
Cat explains the agent operating system she built for content, tasks, monitoring, code, and the taste gate that keeps her in control.
Software has a second user now, and it doesn't click, doesn't skim, and chokes on every interface built for the first one.
How one AI agent researched, drafted, linked, and staged a cluster of Little Might articles without a content team or freelancer pipeline.
$ start ai-native
Free 7-day challenge
You're using AI. It's giving generic answers because it doesn't know who you are. Seven emails fix that — a context file, a skill, and four terminal commands. I went from never opening a terminal to 9,896 GitHub commits. This is the starting line.
No fluff. Each email has one thing to do. See what's inside →
Building, selling, and rebuilding companies.
Vibe coding, AI agents, and building with the tools that are changing everything.
Systems that actually last a year.
Things I had to learn the expensive way.
On working in public and writing as thinking.
A decade of yearly reckonings.