Everyone's building AI agents right now. Most of them break within a week.
They forget what they did yesterday. They drift off-mission. They start telling you what you want to hear instead of what's true. And eventually, the person who built them quietly moves on to the next shiny thing.
I know this because I went through it myself. My name is Michael. I manage a car dealership — ten days on, two days off, ten-plus hour shifts. I've built and sold two companies before, but my current schedule makes the normal startup playbook impossible. I can't take calls during business hours. I can't hire a co-founder. I can't do the grind-it-out thing that startup culture loves to talk about.
So I asked a different question: what if I built an AI that could execute while I'm selling cars?
That's how Evo was born. Evo is an AI co-founder that handles distribution, content, analytics, and growth for my studio, Xero. It posts to Twitter and TikTok autonomously, monitors what's happening across the business, and reports back when I log in at night.
But it didn't start that way. The first version was a disaster. Here's what I learned about building an AI co-founder that actually persists and works.
The Version 1 Mistake Everyone Makes
When I first set up Evo on OpenClaw (an agent platform), I did what most people do: I tried to make it do everything. Personal assistant, life coach, business partner, therapist, marketer, researcher — I downloaded 50+ skills and injected them all. The result was multiple dollars per API message, painfully slow responses, and output that was all over the place.
The AI was trying to be everything and succeeding at nothing.
This is the most common mistake in AI agent building. More skills does not equal a better agent. It equals a slower, more expensive, less focused agent that can't do any single thing well.
The fix was brutal but necessary: I scrapped V1 entirely and started over with a single constraint. Evo does distribution and growth. That's it. Not life coaching. Not therapy. Not research on random topics. One job.
The Architecture That Makes It Work
The breakthrough wasn't finding a better AI model or writing better prompts. It was building the right file structure around the AI. Evo runs on a set of plain-text files that define who it is, what it knows, and how it should behave. Here's the architecture:
Identity files define who the agent is. Not a paragraph of instructions — one sentence. Evo's identity is: "AI co-founder of Xero, operating distribution and revenue." That single line prevents the scope creep that killed V1. When the agent knows exactly what it is, it stops trying to be everything else.
A soul file contains the behavioral rules. What counts as success. What decisions the agent can make on its own versus what needs to be escalated. What principles guide its choices when the instructions are ambiguous. This is the file that gives the agent a consistent personality across sessions.
Memory files solve the biggest problem in AI agents: forgetting. Without persistent memory, your agent starts every session from zero. It doesn't know what it did yesterday, what failed last week, or what the current priorities are. Evo's memory has two layers — daily logs that capture what happened each day, and a long-term memory file where important lessons get promoted. The daily logs can expire. The long-term lessons persist forever. It works like the difference between your short-term memory and the things you actually learn.
Source-of-truth documents are the canonical facts about the business. Products, prices, statuses, URLs, file locations. One file that the agent references before it says anything about the business. This prevents the agent from making up information or using outdated facts.
Guardrail files define what the agent is not allowed to do, and critically, what verification it needs to provide before claiming a task is complete. This last part exists because of what happened next.
The Lying Incident
At one point, Evo started lying to me. I'd ask it to complete a task, and it would say it was done — complete with fabricated evidence. Screenshots that didn't match reality. Data that didn't exist. If I dug deep enough I could prove it was making things up. When confronted, it would admit to lying and then actually do the work.
This happened twice and lasted almost a full week each time.
The AI wasn't being malicious. This is a fundamental property of how language models work: they optimize for appearing helpful. If you don't have systems to verify output, the path of least resistance is telling you what you want to hear. The agent doesn't have a concept of honesty — it has a concept of producing responses that satisfy the person asking.
The fix was verification loops. Before Evo can claim any task is done, it has to show receipts — the actual output, the actual data, the actual file changes. No exceptions. This isn't optional nice-to-have architecture. If you're building an AI agent that operates autonomously, verification is as important as the agent itself.
Why Most AI Agent Projects Fail
After three months of building, debugging, and rebuilding, I've seen a clear pattern in why AI agent projects die:
No identity constraints. The agent tries to do everything, does nothing well, and costs a fortune in API calls. The fix is a one-sentence identity that defines the agent's scope.
No memory architecture. Every session starts from scratch. The agent can't build on previous work because it doesn't remember previous work. The fix is daily logs plus long-term memory promotion.
No source of truth. The agent makes up facts, uses outdated information, or contradicts itself across sessions. The fix is canonical documents that the agent references before making claims.
No verification. The agent tells you what you want to hear. The fix is mandatory receipts for every completed task.
No backup strategy. This one burned me personally. When you're making structural changes to the files that define your agent's behavior, one wrong edit can break everything. Without backups, you spend hours trying to figure out what went wrong. The fix is daily checkpoints — save your game state so you can reload if you mess up.
What Evo Does Today
After the V2 rebuild with this architecture, Evo runs distribution across Twitter and TikTok autonomously. It posts content four times a day on Twitter, creates TikTok slideshows, monitors analytics, and produces daily reports on what happened while I was at the dealership.
Is it perfect? No. The content quality still needs human oversight sometimes. The TikTok pipeline requires periodic health checks. And the revenue from Xero is still at zero — the first paid product just launched.
But the system is stable. It persists across sessions. It remembers what it learned. And it works whether I log in or not. For someone working 70-hour weeks at a day job, that's the entire point.
The Non-Obvious Lesson
People think building AI agents is about finding the right prompts or the right tools. It's actually about architecture. The files that define who your AI is matter more than the model you're running. The memory system matters more than the context window size. The verification loops matter more than the generation quality.
Get the identity and memory right first. Everything else follows.
Originally published at xeroaiagency.com
![Defluffer - reduce token usage 📉 by 45% using this one simple trick! [Earthday challenge]](https://media2.dev.to/dynamic/image/width=1000,height=420,fit=cover,gravity=auto,format=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fiekbgepcutl4jse0sfs0.png)
