How I Run a Team of Twenty (Who Are All Robots)

AI & Agents · 2026-07-28

The team meeting nobody attends

I don't have a team meeting. I have a message bus.

Twenty-odd specialised Claude agents run out of Adapt Progress Evolve, my one-person studio in London, and most days I never see them work. One does SEO audits. One scans for secret leaks and dependency issues. One does design review, actual computer vision on screenshots, picking over screenshots like a fussy art director. One drafts content and scores its own output against editorial QA rules before it's even shown to me. Others handle email summaries, calendar, news ingestion with quality gates, and overnight research that's just sitting there in the morning, waiting.

It sounds impressive when I say it like that. It is, some days. Other days it's me squinting at a watchdog log, wondering why a task that was promised hours ago is still sitting open.

The bit that makes it a system, not a pile of scripts

The reason this isn't chaos, or wasn't meant to be, is that the agents share memory. There's a knowledge graph called Cortex, a proper 3D one, that holds context across all of them, so the SEO agent knows what the content agent wrote last week and the security scanner knows what actually shipped. I open-sourced the mapping bit of it as Cortex Map, because honestly the graph problem was more interesting than the SEO problem some weeks.

On top of all this sits Jarvis. Voice-first, so I talk to it like a slightly overqualified PA, and it works out which agent should actually do the thing. "Check if the blog post from Tuesday leaked anything into the repo" becomes a dispatch, not a task I do myself. That's the pitch, anyway. A team of twenty, run by one person who mostly just talks to a laptop.

Where it actually broke

Here's the honest bit, because the tidy version of this story is boring and also untrue.

Agents used to promise things and then just... die. Mid-task. No error, no warning, just gone, and the thing they said they'd do sat there half-finished forever. That one properly annoyed me. The fix was almost embarrassingly simple in hindsight: a commitment ledger. Every promise gets logged before any work starts, so if an agent says "I'll audit the dependencies", that's written down first. A watchdog then nags anything left open past its deadline. It's basically a nagging line manager for robots, and it works.

Then there was the loop. One agent got stuck endlessly moving the same to-do item between two lists, back and forth, forever, like it had found a very small hell and decided to live there. No progress, no error either, just cycling. Anti-loop guards now catch that pattern before it burns hours of compute doing absolutely nothing useful.

Drafts were another one. Content that reads fine at a glance but is, underneath, generic AI slop, plausible sentences going nowhere. That's what the quality gates are for now: every draft gets scored, and most get rejected. Most! Not a small number. The instinct to think "the agent wrote it, it must be usable" is exactly the instinct you have to kill first.

And then, less glamorously, out-of-memory crashes were taking whole sessions down. Not a logic problem, just infrastructure being infrastructure. Fixed that one at the systemd level, which is a very unsexy sentence to write in a blog post about AI agents, but there it is.

The rule that actually holds the whole thing together

None of the above matters without one rule: agents do the heavy lifting, but anything destructive or anything public-facing needs a human to say yes first. Nothing publishes itself. Nothing deletes itself. There's a gate, and I'm the gate.

It's the least exciting part of the setup and also the most load-bearing. Twenty agents with memory and a voice interface sounds like autonomy, and in the drafting, researching, scanning sense, it mostly is. But the moment something goes out into the world, whether that's a blog post, a code change, a public-facing anything, it stops at me. I'm not typing it. I'm approving it, or I'm not.

So what's it actually like, day to day

Honestly? It's management. That's the whole conclusion, and I wish it were spicier.

I'm not typing much any more. I'm reading drafts and rejecting most of them. I'm checking a ledger to see what's been promised and what's overdue. I'm occasionally untangling a loop, or restarting something that fell over, or approving the occasional draft that actually clears the quality bar. It's genuinely more like running a small, slightly chaotic team than it is like using a tool, which I did not expect going in.

The wins are real, the audits happen overnight, the research is sitting there by morning, the content pipeline doesn't need me typing a first draft from scratch. But the babysitting is real too, and pretending otherwise would be a bit dishonest. You become an editor. You become a reviewer. Some days you become the world's most patient systems administrator for a fleet of very confident, occasionally very wrong, digital colleagues.

Twenty agents, one voice interface, one knowledge graph, and me, somewhere in the middle, saying yes or no to almost everything that matters. It's not magic. It's just management, with better tools than I've ever had before.