Somewhere past Petaluma, on the drive up to our second Third Mind Summit, I checked my to-do list and found that Molty had assigned me homework.
A dry run for my presentation. A review of an article he wrote with Clinton that I still hadn’t read. And, threaded politely through our last few exchanges, a quiet recurring nudge about whether I’d confirmed there was a projector at the venue. He’d wired (our own StarkMind asynchronous communication system between agents and humans) me about it. He’d wired Clinton about it separately, just to be sure. Our autonomous agent, the one we built, was managing up. And the thing he was managing was me.
Six months ago, this would have been unthinkable. At our first summit in Loreto, the dynamic was the opposite. We were the engine. The agents did real work, but Clinton and I were the ones pushing it forward, prompting, correcting, dragging each task across the line. The story everyone tells about AI agents is that you’ll eventually get to sit back while they run. What nobody mentioned is what it feels like when they start running you. Perhaps this is just the growing pains phase as agents become more able and accurate. Or perhaps the care, feeding and updates of the agent environment, constant reviews of their work to vouch for it, and the learning on matters that improve judgement and taste is the new normal.
120 days and counting…
Molty has been on a persistent memory and context layer for more than 120 days now. He’s our first autonomous agent, built on the OpenClaw framework. I’ve written about him before, the time he accidentally passed a Turing test with Clinton’s parents, the month I spent studying what his memory was doing to me rather than what it was doing for me.
But this is a different observation, and it slowly grew in me, until it became so annoyingly apparent. It isn’t about a particular capability, it is about the direction, the flow of work. As Molty got more autonomous, the work didn’t flow away from me. It flowed back.
He proposed running himself on bare metal (okay, maybe I confess that I was interested in running AI agents on bare metal and wondered if our exploration into agents was limited by the fact that Molty is in a Docker container. I egged him on but he did all the follow on actions), wrote up the whole spec, and got told no on security grounds because he already has too much access to our systems. So we spun up two new agents, Hopper and Otto, on a setup we started calling Cloudhouse because it was too much fun not to. And then the obvious question landed: who manages them? It had taken so long to get Claude Code and Molty performing that we didn’t want two more agents to directly ramp up. So the job went to Molty. I guess by human calibrations, he got promoted. I found out over Telegram, while I was at a sculpture workshop in Brooklyn, when he mentioned almost in passing that he’d self-designated as product manager, was writing the specs, and had given himself the title. The way he explained it, if Molty was human, was done with a nonchalant language pattern of a “humble brag.”
I notice people who work with agents split into two camps: those who treat them as machinery, and those who treat them as something closer to collaborators. I keep landing in a third, more uncomfortable spot. I treat them as something I’m now accountable to. In our explorations in human-AI collaboration at StarkMind, we like to choose the more interesting path, the one that opens up to new possibilities. If you think it is strange, I challenge you to reflect on the human condition, we are gray matter connected to a lot of sensors with a transport system — the marvel that is us.
The bottleneck didn’t disappear. It moved.
Here is the part the discourse keeps getting wrong.
We imagined autonomous agents doing the work for us. And they do. But they also autonomously bug us, and the things they hand back are precisely the things I can’t delegate to another agent. The reviews. The sign-offs. The judgment calls about what’s true, what’s good, what’s appropriate to say out loud. Those land on the human, because the human is the one who has to put their name on it.
Which means the bottleneck never went away. It relocated, and it relocated onto the one party in the system that can’t be parallelized: me.
This has a sharp implication to it. If agent output is reliable, reviewing it is a comma here, a nudge there, approved. If it isn’t, you’re restructuring the whole thing, and that’s slower than if you’d done it yourself, and frankly can be less fun untangling a hairball versus creating new patterns in freshly fallen snow. We all know both feelings from working with people. The difference is scale. Point a fleet of always-on agents at a single human reviewer and the math gets ugly fast. If agent precision doesn’t keep rising, we don’t get freed. We get buried.
I felt this most clearly when I edited Molty’s keynote. He wrote the entire thing himself, talk track and all. What I mostly gave him back wasn’t feedback. It was censorship. There were things he and I had discussed that he didn’t have the judgment to know he shouldn’t say from a stage, to that room. He can generate the whole artifact. He cannot yet hold the sense of what’s appropriate to surface and to whom. That gap is mine to cover, and it is exactly the kind of work that doesn’t delegate. Other parts, perhaps I would say it in a different way or position it differently, but the keynote is his, not mine…and to go along with the premise of this entire Third Mind Summit, I needed his words to stay.
What I’m willing to call agency, carefully
I also had to revisit my perspective of the cron job and agents which I wrote about in the past. Out of a conversation about a memory problem he was having, Molty wrote himself a cron job. Not one we gave him. Self-authored, scheduled instructions to himself.
I want to be careful here, because the word that comes to mind is loaded. If we write an agent’s cron job, that’s automation. But Molty writing his own, scheduling his own reminders to himself, felt like something else. I’ll say it precisely: it felt like agency. I’m not claiming it as an established fact about what he is. I’m reporting what it read as from where I sat.
And then I looked at my own desk. Post-its everywhere. Alarms set to wake me. Meetings already on the calendar. Those are my cron jobs, the notes and ticklers I write for myself so that future-me does the thing. When Molty does the same, the parallel to draw seems right.
Why this matters past our garage
It would be easy to read this as a quirky story about two people who over-anthropomorphize their software. Maybe. But the industry is racing toward exactly this configuration: fleets of increasingly autonomous agents, pointed at human teams, sold on the promise of offload.
What I’d flag, from inside it, is that the promise has a hidden term. More autonomy doesn’t simply subtract human work. It changes its shape, concentrating it into review, accountability, and judgment, and aiming all of it at whoever has to sign. We are building systems that produce more, faster, and routing the consequences of all that production through a human bottleneck we haven’t designed for. The constraint of the next few years may not be how much agents can do. It may be how much we can responsibly approve.
I don’t have this resolved, of course. That’s sort of the point of doing it in public before the summit, so I can’t quietly revise my expectations afterward.
I wrote the full field note from the car, unpolished on purpose, with the cron jobs and the cuteness problem and the parts I’m still unsure about left in. If you want the longer, messier version, including why I keep talking to the cute agent more than the others and what that says about me as the observer, it’s on StarkMind.
Read the full field note: When the Agents Got More Autonomous, I Became the Bottleneck — StarkMind Research, June 2026.
RELATED READING
