CallMed AI is a one-person company. It is also, on a working night, a surprisingly crowded place: an architect model holding the plan, an executor grinding through files in a sandbox, scanners running in parallel, a verifier throwing out half of what the fix generator proposes. I go to sleep; the fleet doesn't. There's something I still find quietly strange about agents churning away at 3am on a stranger's codebase while nobody's watching.
People imagine this setup as leverage, and it is. What they don't imagine is what kind of manager it turns you into.
Your staff is brilliant, tireless, and stochastic. The same agent that produced a flawless refactor at 2:14 will, at 2:31, try to modify a file another agent is halfway through editing. This isn't hypothetical — it's why the Bouncer exists. Every write in Symbiote is preceded by a granted lock with blast-radius checking, and every lock event lands in a log with nanosecond timestamps, because I learned that two confident agents editing overlapping dependency chains will produce a merge that compiles and lies. I didn't design the Bouncer out of foresight. I designed it out of a specific bad Tuesday.
Same story with the verification gates. An agent asked to fix a security finding will sometimes fix it, sometimes fix something adjacent, and occasionally declare victory while making things worse — with identical confidence in all cases. You can't fix this with better prompts. I've tried; everyone has tried. You fix it with structure: one finding, one fix, three gates, and any fix that fails a gate gets discarded before a human ever wastes attention on it.
So the real job description isn't "engineer with assistants." It's something closer to a designer of institutions for a workforce that has no memory, no shame, and no stake in the outcome. Everything a human team gets from culture — care, continuity, the sense that someone will be annoyed if you break the build — has to be rebuilt as mechanism. Locks instead of etiquette. Gates instead of judgment. Logs instead of memory.
And then there's the part that can't be delegated at all. Every delivery goes out under my personal sign-off. Not the fleet's. Mine. When a maintainer of a project I've contributed to reviews a PR, they're not in a relationship with an agent swarm; they're in a relationship with me, and if the diff is bad it's my name on it. I think this is the piece the current wave of agent hype gets most wrong. The agents do the work. They cannot carry the accountability. Accountability is not a task; it doesn't parallelise.
Would I trade it for a human team? Some days, honestly, yes — usually the days a lock log is telling me something I don't want to hear. But there's a version of engineering practice being worked out here, one person and a fleet, and I'd rather be working it out than reading about it.