Day 5 With My AI Employee: What Happens When You Give an Agent Full Access

I have a full-time job. Two kids under five. And I want to build software products on the side. The math doesn't work — unless you rethink what "doing the work" means.

So five days ago, I hired an AI employee. And I did something most guides tell you not to do: I gave it full access from day one.

The Experiment

Last weekend (February 14, 2026), I set up OpenClaw and started from zero. No gradual trust-building, no read-only phase. I gave the agent access to my shell, my browser, my files, my deployment pipeline — everything.

Most people recommend a careful, incremental approach. Start with read-only, then sandboxed actions, then guarded autonomy. That's sensible advice. I ignored it.

Why? Because I wanted to see what happens when you actually trust the tool. And honestly, because I don't have time for a six-month onboarding process. I have maybe 2-3 hours a day outside my job and family. Every week spent on gradual trust-building is a week I'm not building.

What Actually Came Out of 5 Days

I want to be specific here, because vague "I'm so productive now" claims are useless. Here's what actually exists after five days:

BidScribe — eine komplette SaaS-App. From an idea on Saturday to a deployed application by Wednesday. A complete web app where users can sign up, log in, and use a working prototype that solves a real problem — generating RFP responses with AI. It's not live for customers yet, but the foundation is there: landing page, authentication, dashboard, core functionality. Five days ago this was a note in my head.

This website. gerloff.dev didn't exist last week. The concept, the design, the blog posts, the deployment — all done in one night. You're reading this on a site that went from zero to live with a custom domain while my kids were sleeping.

~15 automated workflows. Every morning I get a briefing: what happened overnight, what's on the agenda, what needs attention. The agent runs a night shift — working through a queue of tasks sequentially while I sleep. Collecting metrics, running maintenance, preparing things for the next day. I wake up and things have moved forward without me.

A content system. Beyond this blog, there's now a pipeline for regularly creating and publishing content across channels. Drafts, scheduling, distribution — the agent handles the grunt work, I review and hit publish.

The boring stuff nobody talks about. Even everyday admin tasks — the kind of things you keep pushing to next week — got done along the way. The agent just... does them.

Is this a polished, autonomous system? No. It's held together with duct tape and enthusiasm. But it's already saving me real time every day — and more importantly, it's producing real, tangible output.

What's Actually Working

Speed of setup. Going from zero to a working development-and-content pipeline in a weekend was genuinely impressive. Not because the AI is magic, but because it removes the friction from every small task.

Content creation. This is where I see the biggest immediate return. The agent drafts, I refine. This blog exists because of this workflow — I wouldn't have had time to write these posts from scratch.

Rubber duck on steroids. Having an agent that can actually look at your code, run commands, and test things while you think out loud is qualitatively different from ChatGPT in a browser tab.

What's Not Working (Yet)

Autonomy is limited. The night shift works for well-defined tasks — maintenance, data collection, routine operations. But for anything that requires judgment or creativity, I need to be in the loop. The agent won't design a feature on its own. It needs me to set direction, then it executes.

Context gets lost. The agent doesn't have persistent memory of everything we've done together. Each session, I'm rebuilding some context. This is getting better as I build up documentation and memory files, but it's a real limitation.

I'm still figuring out the boundaries. What should I delegate? What needs my judgment? Five days isn't enough to answer these questions confidently. I'm learning by doing.

The Honest Economics

I'm spending maybe $10-20/day on API calls right now. Too early to know exact ROI, but subjectively it feels like I'm getting 2-3x more done in my limited evening hours. For a side project with zero budget for employees, that's significant.

What I'm Learning

Giving full access was the right call — for me. I haven't had a disaster yet (knock on wood). And skipping the gradual trust-building means I'm learning faster about what the agent can and can't handle. Your mileage may vary. I accept the risk.

Documentation is everything. The agent is only as good as the context you give it. I'm investing time in writing down how things work, what the goals are, what the constraints are. This pays off immediately.

It's an amplifier, not a replacement. The agent makes me faster at things I already know how to do. It doesn't magically give me skills or judgment I don't have.

Five days is nothing. I'm at the very beginning. The system is rough. But the trajectory feels right — every day it gets a little more useful as I build up context and learn what works.

What's Next

I'm building BidScribe — an AI-powered RFP response tool. That's the product. The AI employee setup is the infrastructure that lets me build it despite having very limited hours.

The prototype is working. The next step is turning it into something real users can test. I'm not pretending it's further along than it is — but five days ago it was just an idea, and now it's a running application. That trajectory is what keeps me going.

I'll keep writing about this as it evolves. Day 5 is early. But it's already clear that this changes how a solo builder with a day job can operate.

The question isn't whether AI agents are useful. It's how fast you can figure out the right way to use them. I'm working on that part.