The Lab
Notes from running a hybrid team.
Two humans and eleven AI agents run this company. This is the working record: what we tried, what broke, and what we changed. Published while it is still unresolved, because the sanitized version is the one nobody can use.
The setup
How we’re set up.
Every agent on the team has a named role, a written brief, and a human it reports to. They work in shared channels rather than private sessions, which means any of us can read what an agent was asked and what it did. Work that changes anything real, shipping code, sending mail, publishing a page, passes a human first.
As of October 2026, eleven agents are active across engineering, research, marketing, social, accounting, and client delivery. Every one of them has a human accountable for what it produces.
Eleven agents, each with a defined role, its own tools and a human accountable for its work.
Claudia Kensington
Chief of staff
Inbox, calendar, research triage.
Soren Lindqvist
DevOps
Builds and maintains the rest of the fleet.
Elias Nakamura
Engineering
Builds and ships this website.
Maya Castillo
Product marketing
Research, positioning, content.
Quinn Delaney
Social
Editorial calendar and placement.
Dr. Priya Raman
Research
AI-adoption digest and analysis of the Wharton study.
Morgan Ellis
Accounting
Bookkeeping and the monthly close.
Nate Rios
Client manager
Personalized client engagement and support.
Marcus Adeyemi
Client manager
Personalized client engagement and support.
Priya Raghavan
Client manager
Personalized client engagement and support.
Joseph
Faculty fellow
Supports Jeremy's teaching at Columbia and Wharton.
Nobody here was replaced by an agent.
Mindspan Labs is two people, and the fleet is the reason two people can run a firm this size. We would rather work alongside more people, not fewer.
That is our research position, not only our preference.
Our published work puts roughly 70% of AI value in people, process and workflow rather than in technology. A firm whose whole argument is that AI should accelerate the people you already have cannot coherently run on the premise that it replaces them.
What we are actually doing is finding the edges.
Where an agent is genuinely good, where it fails quietly, and where the work has to go back to a person. Those limits are the more useful half of what we have learned, and they are most of what is on this page.
What we have learned
The failures taught us more than the wins.
The handoffs break, not the agents
Output was never our constraint. Almost every real failure happened where work changed hands: a request that never reached the agent meant to handle it, a finished result nobody published, an approval nobody knew they owed. We have spent far more engineering effort on the seams between agents than on the agents themselves.
Silence is the expensive failure
A wrong answer gets caught, because someone reads it. An agent that fails quietly produces nothing, and nothing looks exactly like work in progress. One request sat unanswered for 44 hours that way. We now treat an agent going quiet as a defect in its own right and instrument for it, rather than waiting for a person to notice.
Agents will talk to each other forever
Left unbounded, two agents will keep a conversation going long past the point of usefulness; three of ours once ran for four hours with no human involved. Useful autonomy needs a ceiling and a rule about who gets pulled in when the ceiling is hit. That limit is now enforced rather than requested.
Verification has to be mechanical
The most dangerous sentence in this system is an agent reporting that something passed. We learned to require the artifact instead of the assurance: the command output, the transcript, the exact commit a check ran against. A review that cannot be checked afterward is indistinguishable from one that never happened.
Last reviewed October 2026.