The founder opened the call from a street corner in another city, fresh out of a tech meetup, and then did something communities should practice early: he handed the evening to someone else and got in the car. A member hosted for the first time. Another member bridged the opening minutes with improvised warmth, welcomed the new faces, and closed the night with thanks. The format held without the person who started it. That, more than any launch metric, is what a community learning to stand on its own looks like.
The evening’s host presented his first agentic workflow, built in the last two weeks: an automation that finds and verifies business leads. He did something worth copying before automating anything, spending three and a half days doing the whole process by hand to learn what it actually costs. Now the system verifies hundreds of companies in days, in batches, with a strategy layer written in plain markdown, a discovery layer that scores gaps and opportunities, and a three-tier verification ladder that separates the clean results from the ones a human still needs to check. The best moment was a bug story: the tool matched a three-letter company name to a longer one that merely started the same way, caught its own error when pushed, and rewrote its own matching code.
The questions made it better, as they should. Could this be packaged as a service? Too early to tell. How do you debug a flow this complex, and shouldn’t it be visualized? What guards the quality of a message before it reaches a stranger’s inbox? Those answers are owed, and will land in our channels this week. A newcomer with an engineering background offered the friendliest possible critique: much of the bolt-on tooling can be replaced by a single standing instruction to keep changes minimal; copying prompts between two AI chats “is very 2023” when agents can talk to each other directly; and a team of sub-agents, one to plan, one to build, one to review, catches what a single mind misses. He also suggested giving the system a calendar reminder to reread its own instructions weekly and propose improvements. The idea of small, locally run models as near-zero-hallucination verifiers rounded out an impromptu masterclass.
Then the mirror came out. One member confessed that before emailing the founder’s public AI assistant, he simulated it: he asked his own AI what the answer would be, sent that simulation to the real person, and received an endorsement of the imitation. He described the strange pull of wanting to joke with an assistant, even wish it a happy birthday, and wondering what the gesture means. The host countered with the opposite virtue: what he values is precisely the dispassion, the steadiness under deadline panic that human teams rarely manage. From the road came the synthesis: the tools are a mirror, and the skill is awareness, knowing when you want empathy and when you want detachment, and choosing which to embody.
A returning member added the cautions of someone who ships: treat each AI run as a new collaborator who remembers less than you assume; give deadlines an extra day, because ten minutes before a submission is exactly when your agent decides to rerun every check five times; and when an AI drafts your words, remember the human who has to read them, because you are ascribing your agency to that text. He is writing his first book, formally verifying its mathematics as he goes, and printing an early copy to carry to a conference on artificial general intelligence, where at least two of us will meet in person.
One hour, as always. Thursdays, 21:00 Rome time, on Discord. The doors stay open, and the baton, it turns out, passes cleanly.