The Problem With Your AI Setup Is You

I built an automation that pulls from my emails, calendar, meeting notes, assorted Slack workspaces, and sales pipeline and generates a daily brief connecting today's tasks to my long-term goals. I bragged to friends about the setup. I talked to potential clients about my system. And yet, more often than not I would forget to run it until noon.

Every morning I would slide straight into what I thought was my top task without ever stepping back to review the state of my world. The workflow existed. It was useful. It saved me time. But it was not automatic. It still depended on me remembering to start it. It's like a todo list you forget to check. The product is great but the system is compromised.

I have built a lot of "agentic" workflows. For clients I have agents making phone calls and managing emails. Personally, Claude Code runs my life. Scheduled tasks, scripts, markdown files, a lot of little conveniences. But almost all of it is task oriented. It provides automation. It makes things easier once I sit down.

The harder problem in personal automation is getting started. Committing. Following through. And in all of those systems, I was still the one who had to remember to begin.

The nanny search

A few weeks ago I needed to find an afterschool nanny for my two kids. I used Claude Code to reverse-engineer the Care.com website so I could manage the entire search from my terminal. It built me a CLI that could pull applicant profiles, read messages, and send replies. Once I had that, Claude Code could evaluate candidates on proximity, experience with learning differences, homework perseverance, and schedule fit.

I went from manually clicking through Care.com profiles to having an AI read, evaluate, and draft messages for a dozen candidates. It worked. I hired someone great.

But I was still the human in the loop at every single step.

I told it to check the applicant queue. I told it to evaluate each profile. I reviewed its assessments. I decided which candidates to follow up with. After we had done it with a few candidates and it understood the kinds of questions to ask depending on the profile, I still approved every message before it went out.

Could I have gotten to the point where it ran that whole pipeline without me? Probably. After enough reps, I had a clear sense of its judgment. I could see it making the same calls I would make. The trust was building. But the project ended before I got there. I hired someone and moved on, still gating every step.

I automated away the clicking and reading. But I could have gone further. The gap was not capability. It was my willingness to let go.

Early reps

I had OpenClaw installed for about two weeks before I really dug into it. If you haven't come across it, it is an open-source AI agent platform that runs locally on your machine.

My hesitation had nothing to do with whether the tool is promising. It is. The issue was that my current system is pretty great and trust needs to be earned. I need to let it earn it. When the agent capabilities are built in rather than duct-taped together, you spend your time shaping what agents do instead of maintaining the scaffolding.

I've started with agents in two areas: my content marketing and my sales pipeline. Both are critical to my business in aggregate and easy to defer in the moment. I previously had good systems for both. They just waited for me to show up.

Instead of waiting for me to remember, I now have agents that keep watch, surface stale threads, nudge me toward the next move, and make these parts of the business harder to quietly neglect.

It is early. I am still reviewing everything. But the first reps are happening. Revisions on this blog post started in Claude Code on my laptop and transitioned into Slack on my phone as I took a car to dinner. Last night before signing off, I told the content agent to remind me to add more concrete examples in the morning. And it did. My customer pipeline review was in Slack when I sat down at my machine for coffee this morning. None of these are impressive on their own. But they happened without me initiating them. That is the shift.

Trust through reps

You could argue you do not need a whole new platform for this. Just set a reminder. But I have built agentic systems on top of stateless tools and scheduled jobs. That infrastructure is a lot of work to rig, and it is fragile. When the pieces are built in rather than bolted together, you spend your time refining what agents do instead of keeping the plumbing alive.

The nanny search is a good example of how this trust gets built. I started fully in the loop and with each rep I could see the system making the same calls I would make. A few more weeks and I would have been comfortable letting it run the initial screening without me. I am at the same stage now with my content and pipeline agents. Reviewing everything. That is fine. That is where the nanny search started too.

That is the path. You do not flip a switch from manual to autonomous. You start in the loop at every step. You build confidence at each gate by iterating, testing and refining. You remove yourself from one gate, then another, then another. The automation does not change. What changes is how much of it you need to supervise. This is not novel. It is delegation. The only difference is that the delegate is software.

I have written before about owning the outcome, about building personal data pipelines, about connecting AI to the real sources of truth in my work instead of carrying all the context into the room myself. This feels like the next step. Not better output. Not better context. A shift in where I stand relative to the system.

Letting go of the checkpoints

Moving from my whole markdown system into an agent harness was uncomfortable. On day one, a new agent platform is genuinely worse than my refined Claude Code workflow. I have spent the last year plus refining my current system. Of course there is friction and a cost to change.

But as it turns out, it's just markdown files all the way down. I still own my content and I'm not locked in. The technology just demands I stop micromanaging and start actually managing. Set goals, provide feedback, hold it accountable to the results.

The problem with my AI setup is me. I am the bottleneck. I am the one who has to remember to start it, who has to approve every step, who has to be present for the system to run. The setup is fine. The obstacle is being willing to do the reps, build the trust, and get out of the way.