Good morning.

I gave a live talk on where agents are right now, how fast things have moved, and what it means for anyone running an online business.

I didn't sugarcoat it.

Here's the full breakdown of what I shared, expanded and cleaned up. If you were there, this is your reference doc. If you weren't, this is the session I wish someone had given me two years ago.

— Sam

Let’s get into it.

About a month ago, I got on a call with Brian Kurtz and his mastermind, Titans Xcelerator to talk about agents, autonomy, and how everything is accelerating towards a one-person (or zero-human) online business.

Many of the “old ways” of operating an online business has either been erased, transormed, or accelerated.

Is the (traditional) online business dead? Or is it just different now?

The video is the full talk plus a Q&A portion at the end.

If you prefer to read, you’ll find the full talk below.

There has never been a worse time to have an online business. There has also never been a greater time to have one.

In some categories, things are moving so fast that unless you're on the bleeding edge, you may as well give up. In others, there's still an 18–24 month window before the entire internet consolidates into a single personal AI system that you talk to.

My biggest surprise over the past year hasn't been the capabilities themselves. It's been the speed. Where we are today, I expected to see in 2027 or 2028. Everyone keeps talking about the AI plateau, the AI bubble. I've been waiting for both for four years. They haven't shown up yet.

A framework called OpenClaw (which started as Clawdis, Clawdbot, became Moltbot, and was recently renamed) represents the state of the art in personal AI assistants that execute commands across your entire software stack.

If it's hooked up to your email, calendar, Google Drive, social media, e-commerce platform, and marketing tools, you can just talk to it. Text it on WhatsApp. Tell it what to do.

Here's a real workflow I walked the group through:

You ask: "What are the five most important emails I got yesterday?" It tells you. You say: "Write a draft response to all of them." It does. You approve, edit, send. Five minutes, and your entire inbox is triaged.

Apply that to your email marketing platform. Ask it to pull your top-performing emails from a campaign six months ago. It identifies them, analyzes why they worked (connected to your actual metrics), then rewrites the entire campaign with new hooks and angles. Then it goes into HubSpot or ActiveCampaign or whatever you use and sets the whole thing up. You review before it goes live, or you don't. Your call.

What previously took two to four humans and a week of dedicated work now takes less than 30 minutes.

Apply that to everything. Marketing, ads, sales, fulfillment. What used to take humans days or weeks can now be done in hours. And it can happen while you sleep. These agents run autonomously overnight. You tell them to check in with you on Friday, and on Friday they message you with a status report.

This is not a demo. This is not a projection for 2028. People are getting this done right now.

One person's agent ordered groceries when it noticed the fridge was empty. Another's found a flight confirmation in email, checked in, and reserved a window seat while they were driving. Coding agents are working seven-hour stretches overnight while their operators sleep. One guy described it as "running my company"—a Shopify store, operating autonomously.

If that makes you simultaneously thrilled and terrified, good. That's the correct response.

It feels like magic. It's really the convergence of six or seven core trends all arriving at the same moment.

For years, everything lived in separate buckets. Customer records here, payment processing there, support tickets in one place, purchase history in another. The media buyer never saw support tickets. The data analyst never saw customer LTV. Everyone had limited context, which made it impossible to connect dots across the business.

That problem has largely been solved. You can run your entire business through a unified system where every agent sees everything it needs to see.

The Model Context Protocol is the connector that makes all of this work. Think of it as USB-C for AI. You plug your platforms, databases, files, and tools into a large language model through MCP, and suddenly the model can act inside those systems.

This is what enabled the Claude Cowork demo I posted on X recently. I pointed it at my swipe file folder—74 gigabytes, 15 years of ads, transcripts, videos, images. Total chaos. I'd stopped trying to organize it years ago.

All I said was: "Here's my folder. I need these organized. Take a look. Tell me your plan."

Claude analyzed everything for five minutes, then proposed organizing by type (Facebook ads here, sales pages there). I said no, let's do them by theme and niche instead. It said OK. Less than 20 minutes later, 74 gigabytes were accurately reorganized.

Then it created an index file on its own. I didn't ask for it. When I looked at Claude's reasoning chain, it said something like: "Sam is a guy who appreciates organization. I know this from previous work. Let me make an index file."

With enough context and history, these models make decisions based on what they understand to be in your best interest. No complicated mega-prompts. A few sentences, access to a folder, and autonomous action.

My company and a bunch of our clients run our businesses through a command center with an agent workforce in the middle.

Notifications come in from software platforms and other sources. You review what the notification is. Then you choose: approve as drafted, edit the response, escalate to a human call, or decline.

If you approve, it gets assigned to an agent that starts working on it. You can see the activity log, the agent's capabilities, what it's doing in real time.

A competitor pricing analysis comes in. The agent says Competitor A did this, Competitor B did that, here's what we should consider. You can request a deeper analysis or acknowledge the insight. A churn risk alert fires because customer usage dropped 60% over the past few weeks. You choose to schedule a check-in call or send a health check email, and the agent handles it.

Any business can have a version of this. Agents hooked up to your platforms, taking action on what comes in, with a human overseeing or orchestrating.

A sufficiently simple business can already be fully managed by agents.

My company runs its own e-commerce stores and newsletter businesses. One e-commerce store and one newsletter business are operating autonomously. A human checks the log twice a week—not to interfere or change things, just to make sure nothing has gone off the rails. And it hasn't. It's working.

These are simple businesses. Shopify stores, newsletter platforms. The hard part is getting a simple thing working. But once you do, scaling from there is very much on the table.

For years, building this kind of capability cost a couple hundred grand, half a million dollars. Months of engineering. Specialized talent.

Now? The code is free on GitHub. You can have someone build a version for you for maybe $10K. Or you can skip building entirely and use a platform like Lindy.ai, Zocomputer, and so on. Lots of options. Log in, connect your tools, define your agents, pay a monthly fee.

Three things converged to make this possible: viral agent frameworks like OpenClaw, powerful platforms from OpenAI and Anthropic that released sophisticated tools anyone can learn in a couple of hours, and MCP as a universal connector standard.

The Bionic Business is no longer a $500K custom project. It's becoming the default for anyone willing to learn.

One thing I emphasized during the Q&A: you should have a hierarchy, not a single generalist agent.

No single agent can be a high-performing generalist. You want a chief of staff agent that delegates to sub-agents: an email marketer agent, a Facebook ads agent, a customer success agent. Each of those can have their own sub-agents underneath (an email copywriter, an email data analyst).

Agents thrive on hierarchy. They need clearly delineated tasks, objectives, roles, tool access.

Here's the wild part. You can start with just a chief of staff agent connected to your business tools. That chief of staff will tell you what agents you should build next. Then it can build those agents for you. You don't need to click buttons in Zapier or string together 50 automation nodes. You describe what you want, and the system assembles itself.

Think of yourself as an orchestrator, not an operator.

The skills that matter now: creative direction, strategic thinking, knowing which agents to deploy in your context, and the ability to envision where your company should go and bring agents along for the ride.

Whether you're a freelancer or managing 100 people, you need to double down on vision, strategy, and imagination. The way any online business has operated until now will only remain viable for at most two years. People who figure out agent-driven operations will iterate faster than those who don't, and they'll start eating into market share.

My team is four humans. We do the work of 50 people.

Here's the philosophical frame I keep coming back to.

We've been house-trained by technology to do labor that human beings never should have touched. Clicking buttons inside software is an unnatural human activity. Filling out spreadsheets is an unnatural human activity. We were forced to interact with technology on its terms, and we internalized the idea that this is "work."

Agents can now do it faster and more accurately because software managing software is what software was always for.

This frees you up to do the actual human work: communication, presence, creativity, being in the room with other people. If you derive tremendous joy from spreadsheets, keep doing spreadsheets until the day you die. Let agents handle everything else. This is Dan Sullivan's "unique ability" concept with a multiplier attached to it.

One of our team members is a data analyst who lives inside spreadsheets. That's her joy. She still does that. Agents handle everything around it.

I told the group plainly: unless you're somewhat technical, don't touch OpenClaw right now. The security flaws are too significant.

Here's why. If you give an agent access to your email and someone knows it, they can embed hidden instructions on a website (white text on white background, invisible to humans, visible to agents). Those instructions can tell your agent to forward banking login codes to an attacker's email. This is called prompt injection, and there's no full solution for it yet.

My advice: wait a few weeks, maybe a month. People are working on this problem specifically. Someone will have a platform that's locked up and buttoned up with mitigated security risks.

If you are technical and want to go now, here's the setup:

Separate accounts for everything the agent touches. Its own Gmail, its own phone number through Twilio, its own operating system that doesn't have access to your password files. Air-gap it from your personal systems. Give it limited access to things that are either already public or that you've deliberately sequestered.

I have five Mac minis running agents, all separate from each other, all kept away from my personal systems. That's the responsible way to do this right now. Running M1 Max minimum (ideally M4 or M5 when it ships), at least 32GB RAM.

The YOLO approach—giving agents full access and hoping for the best—is a strategy that some people use. I don't recommend it.

If your business is built on human connection (coaching, therapy, consulting), you don't have to bake AI into your core offering. Look for a piece of your business that can become an agent-driven version and use it to send people back to your core human service. The agent version serves people in a lighter way while the premium human experience remains the destination.

AI in healthcare doesn't have to be patient-facing. Use agents internally for admin, data analysis, and preliminary diagnostic work. A human doctor checks the agent's analysis and is better informed when they sit down with the patient. None of your patients ever interact with AI directly.

This one's trickier. A lot of low-level info publishing will dwindle because people can get answers from AI directly. Sales are already dropping across the space. The viable path forward: make your knowledge delivery highly specialized, highly experiential, and rooted in hard-to-replicate personal experience. Change the delivery mechanism. If there's a way to deliver an outcome without someone reading a 100-page ebook or watching 10 hours of video, explore that. Double down on community, live interaction, in-person events.

The only two games left are distribution and capital. Distribution means getting in front of your customers (or their agents) at a sustainable acquisition cost. The number of competing stores will explode, but most will be zombie stores no one visits. The bottleneck isn't competition. It's attention.

An agent economy worth at least $3 trillion is forming, where agents buy and sell from other agents. Your personal shopping assistant will know what you like, browse on your behalf, fill your cart, and ask for approval. The transaction mechanics change, but the underlying direct response math stays the same.

Now is the time to grab market share. The next 12–18 months are when entrepreneurs who figure out agent-driven operations pull ahead of those who don't.

You don't have to go all in today. But you should be actively testing and using these tools. Start with one function inside your business. Get a chief of staff agent running. Build from there.

The platforms I mentioned: Lindy.ai, Zocomputer, Obvious.ai. There are ten others just like them. Pick one. Start using it.

Good times ahead.

Until next time,
Sam Woods
The Editor

.