When I started building firmd, I knew I needed a real test case. Mechanics were the easy half. The harder question was whether an agentic firm could produce something genuinely useful. I needed a second idea to seed into a firmd pilot tenant and watch it run.
debunkd.social became that second idea.
A few parents I talked to told me I should drop the agentic firm and work on debunkd.social instead. (They had a point. Still doing both.)
debunkd is a (fake - still...) stealth-mode startup with a narrow purpose: help kids at the age they first meet social media learn to spot the dark patterns those apps use to keep them hooked. Endless feeds, streaks, variable rewards, default sharing, asymmetric friction. The plan is simulation apps — safe clones of TikTok, Instagram, Snapchat — that show the mechanism, name it, and let the child decide what to do next time. Not "apps are evil". Pattern recognition.
Then I handed the idea to firmd first customer - the company debunkd - and stepped back.
The Mac Mini that found a buyer
A few days ago I ran a strategy mission on qwen3.5:27b. Mid-tier, open-weights, running locally on my Mac Mini (no GPU :-( ...) .
I had my own ideas about monetization for debunkd. Not kids, not really parents either — my guess was schools, paying a flat fee to support pupils. I never wrote that anywhere, not mentioned to any firmd product manager agent. The only context firmd's debunkd tenant had was roughly what is written above with a few "financial constraints" (pre-seed), ingested into firmd's content management system as available context.
The strategic debate is wanted, by design, that's a core design bet of firmd. I primed the PMM agent on targeting, positioning, messaging, and buyer persona considerations, and instructed to push back when a strategy ships without a credible monetization plan — which mine very much did.
That is where it surfaced something I had not considered: insurances. Insurers carry real litigation exposure when kids and parents are not well-informed about platform harms (the Meta, Epic, and NGL cases are no longer hypothetical). A literacy product reduces that risk. They have an economic motive to pay.
I did not see that one. Clever. From a model running on my Mac Mini. That motivated me to move on...
The first baby step
Yesterday I ran a full cycle end to end on a mid-tier deepseek model. And the next surprise: The past few dozen mission runs the debunkd product crew produced a wall-of-text'ish marketing website for debunkd.social. This time, it was something new ...
This went from strategic scruitiny from multiple lenses, tactical planning and delivery (coding) and back. I had to actually click on it to generate the learning signal to close the PDCA loop for the strategic mission so it updates its prediction error in its causal model.
And it made me smile.
Still a long way to go, but that's the way... The increment is small. There's a lot to be done and I hope for a good ROI on the eval system effort investemt (=next article). But this was the wow moment I needed to share.
The second idea was supposed to be a test case for the first one. Right now they are taking turns surprising each other. ;-)