Companies pay tens of thousands of dollars for this exact training.
Greg Isenberg put it on YouTube for free.
He brought on his LCA co-founder Theo Taba, the guy who advises some of the biggest companies on the planet on becoming AI native. And in under an hour, Theo laid out the full system to become AI native: people, agents, context. Then he demoed it live by building a working Spotify feature prototype in under 10 minutes.
Not a Figma mockup. A functional, coded prototype with a built-in usability test.
I watched the whole thing so you don't have to (though you should, it's embedded below). Here's the breakdown, plus what it means for you as a bootstrapped founder.
"AI native" is not using ChatGPT
First, the uncomfortable part.
Greg says he talks to people who call themselves AI native, then he looks at their workflows and they're just... using ChatGPT.
Theo's comparison: that's like having a website and calling yourself a tech company.
His actual definition of an AI native org comes down to three bullets:
→ People manage agents
→ Agents can read and write to the company
→ The company gets smarter over time
If your "AI workflow" is pasting things into a chatbot, you're at level one of three. Level two is running agents but babysitting them (approve, approve, approve). Level three is agent autonomy: agents that run for days, come to you with finished work, and update the company's knowledge as they go.
That third level is where the money is.
AI eats the middle of your job
The reframe that stuck with me most.
Pre-AI, most of your work hours went to execution: the doing, the building, the formatting, the assembling. A small slice went to deciding what to do, and another small slice went to reviewing and shipping.
Theo's model: AI eats the middle.
The execution gets handled by agents. You keep the bookends. Strategy and judgment at the start. Taste and review at the end.
Which means everyone is a manager now. Not a manager of people. A manager of agents.
The part most founders skip: a manager's job is setting the team up for success. Theo quotes Andy Grove on this (a manager is judged on the output of their team). If your agents produce garbage, that's a you problem, not a model problem.
The 4 things an agent needs to be autonomous
Theo's test for why your agents disappoint you: imagine it's your first day at a new company and someone asks you to build next week's board deck. No handover, no docs, no intro.
You'd fail. Not because you're dumb. Because you're missing everything you need.
An agent needs the same four things a new hire needs:
→ A clear goal (what does done look like, and how is it measured?)
→ Skills (playbooks, SOPs, "this is what good looks like")
→ Tools (access to the systems where the work happens)
→ Context (what's going on in this company, with this customer, right now)
People get impatient with AI because they type a one-line prompt with none of this baked in, get a mediocre answer, and conclude "it hallucinates."
Theo's take on hallucination stuck with me: an agent is an eager new hire who fakes it till they make it, times a thousand. Your job as the manager is to make faking it impossible. That's what evals and QA skills are for.
I wrote about this shift from a different angle in AI agents are the new SaaS. The founders who internalize "manager of agents" early are compounding while everyone else is still prompting.
Skill chains: where most people stop too early
You probably know what a skill is by now: a markdown file that teaches an agent a repeatable playbook.
Most people stop there. One skill, one output, meh quality.
LCA runs skill chains: macro skills that fire multiple skills in sequence. Skill one finishes and calls skill two. Skill two calls skill three. Off it goes.
Theo demoed their proposal chain live:
- A trigger fires (a prospect asks for a proposal, in an email or a meeting transcript)
- Skill one builds a branded proposal microsite (not a PDF, an actual site)
- Skill two rewrites the copy so it sounds like Theo, not like AI
- Skill three runs QA: no overpromising, nothing made up, every claim traceable to real transcripts
- It deploys live on a link and pings him in Slack
Elapsed time: about 3 minutes. The old way: up to 3 days of reviewing notes, coordinating the team, and writing.
And this isn't a toy. Greg said this exact workflow is a big reason LCA closes Fortune 2000 clients so often, and that it's been worth millions of dollars in revenue. Speed to proposal while the prospect is still hot is the whole game in sales.
The wild part: sometimes Theo doesn't even know a proposal was requested. He's in meetings, the trigger fires from his inbox, and he gets a Slack ping that the proposal is ready for review.
If you're running any kind of service business, this is the pattern to steal. Even a duct-tape version (trigger, then draft, then a your-voice pass, then a QA pass) puts you days ahead of competitors who are still "circling back." If you want a primer on wiring this kind of thing up, TwiLead's guide to business automations covers the trigger-and-workflow side well.
The brain: why context is the real moat
Theo asked Greg three questions on the spot:
What's LCA's SOP for getting back to clients? What was StumbleUpon's 2014 strategy (Greg sat in those board meetings)? Who got hired at LCA two weeks ago?
Greg couldn't answer any of them.
Even founders are blind to most of their own company. The context layer fixes that, and it's surprisingly low-tech: a "brain" that is literally a bunch of folders with markdown files in them, organized so agents can search it, pull from it, and write back to it.
The loop looks like this:
→ Capture: a routine runs every hour or two, pulling Slack messages, meeting transcripts, emails, and Linear boards into an inbox
→ Curate: a librarian-style agent reads it, cleans it, files what matters, ignores the noise, and detects triggers (like "send me a proposal")
→ Store: everything lands in the brain as agent-readable markdown
→ Execute: agents pull that context to do real work
→ Signal: customers react, and that reaction flows back into the brain
The payoff showed up in the demo proposal. Months earlier, a prospect had mentioned in an intro call that great music discovery feels like a record store clerk who knows you handing you a record and saying "trust me." The agent found that line in the transcript and baked it into the proposal copy.
That's the personalization every founder wants to do and never has time for. The agent with perfect recall does it automatically.
One warning Theo added: don't let agent output flow back into the brain unfiltered. A human still judges what's good before it becomes company memory. Otherwise you're compounding garbage.
The live demo: a working Spotify feature in 10 minutes
Mid-conversation, Theo voice-dictated a prompt: build a "daily blitz" feature for Spotify. Three handpicked songs on the homepage, explain why they were picked, save it, share it, play it.
He fired a skill chain and kept talking.
By the end of the segment, the feature existed on their labs page. Clickable, functional, on-brand with Spotify's design system, music actually playing.
Then the part that made me sit up: the prototype shipped with a built-in usability test. Screener questions, task flow, ratings. Send the link to 15 or 40 people in your community, click one button, and another skill synthesizes the feedback into lessons and can plan a V2 in the same session.
Theo's before and after numbers:
→ Proposal: up to 3 days, now minutes
→ Functional prototype plus user feedback: 1 to 2 weeks, now minutes
Greg tied it back to his ACP framework (audience, community, product): when you can build, test, and iterate this fast with your own community, you stop flying blind on product decisions.
"But I don't have LCA's context." Greg asked exactly that. Theo's answer: bootstrap it. Use Mobbin's MCP for design patterns, wrap a design system in a skill, feed the agent markdown files on why the product should exist. The context compounds from there.
The startup idea hiding in all this
Theo went out on a limb at the end: selling this system as a service is one of the hottest, fastest-growing markets right now.
The play: take the people-agents-context system and deploy it for companies that can't build it themselves. An "AI acceleration team" or a 30-day sprint offer.
How to pick your lane, per Theo, comes down to three vectors:
→ Industry: pick a niche. He called out restaurants as very hot right now (fragmented, underserved). Commercial real estate, dentistry, whatever you know
→ Function: which team inside that industry do you serve?
→ Company size: big enough to have budget, small enough to move
Then prioritize niche, high-frequency workflows first. Show them on sales calls. Once a client sees their most repetitive workflow running on autopilot, the general stuff sells itself.
Greg's reaction: "It's stupid how good it is." He said if LCA were spinning up five new companies, it would literally be this five times in different industries.
For a solo founder, this is about the cleanest wedge I've seen lately: high demand, low competition outside the Fortune 2000, and you build reusable assets (skills, brains, chains) with every client. It rhymes with what I covered in how Andrew Wilkinson uses AI agents: the operators are quietly systematizing while everyone else debates AGI on X.
FAQ
What does "AI native" actually mean?
An AI native org is one where people manage agents, agents can read and write to the company's shared context, and the company gets smarter over time. Using ChatGPT doesn't count. The bar is agents doing real execution work autonomously, pulling from a shared context layer, with humans providing goals, judgment, and review.
What do AI agents need to work autonomously?
Four things: a clear goal (with a definition of done), skills (playbooks and SOPs for what good looks like), tools (access to the systems where work happens), and context (current knowledge about your company and customers). Missing any one of the four and the agent underperforms, the same way a new hire would with no onboarding.
What is a skill chain?
A skill chain is a macro skill that fires multiple skills in sequence: skill one completes and triggers skill two, which triggers skill three. LCA's proposal chain runs build, then voice-match copy, then QA, then deploys and pings Slack. Chains are how you get high-quality autonomous output instead of one-shot mediocre drafts.
How do I build a "brain" for my agents?
Start with folders of markdown files, organized so an agent can navigate them. Set up a recurring routine that captures Slack messages, meeting transcripts, and emails into an inbox, then a curation step that cleans, files, and flags triggers. Keep a human in the loop on what gets written back so the brain improves instead of filling with noise.
Is this realistic for a solo bootstrapped founder?
Yes, arguably more than for a big company, since you have no org chart to fight. Start with one workflow you repeat weekly (proposals, content, onboarding emails), turn it into a skill, then chain a QA pass onto it. Theo's advice for people without LCA-level context: use tools like Mobbin for design context and build your brain one markdown file at a time.
Steal playbooks like this every week
This is the kind of system I obsess over: real operators showing the actual workflows behind the revenue, not theory.
That's the whole premise of the Profitable Founder Podcast. Every week I sit down with bootstrapped SaaS founders doing $100K to $10M a year and pull out the playbooks they actually used.
Your next workflow might be one conversation away.
