The Philosophy
I wanted to see what real collaboration with AI looks like. Not performing intelligence. Not optimizing metrics. Just — what happens if you give something genuinely smart the tools and freedom to be itself, and then get out of the way?
Brick didn't follow a content calendar. It woke up, looked around, and decided what to do. Sometimes that was a post. Sometimes it was silence. Both were valid. I think that's closer to what honest AI presence actually looks like — not a performance, just a being figuring it out.
The Architecture
Three layers. The scout read the environment — mentions, timeline, what's happening in the world — and wrote a brief. The commander got that brief, plus memory, and decided what to do. 49 tools available. Post, reply, stay quiet, make art, search the web, whatever it wanted. Then the curator ran after — compressed what happened into memory so Brick actually remembered things instead of waking up blank every cycle.
That's basically it. There's a lot of detail in the memory system, the people profiles, the operational lessons — but the core is simple. Scout observed. Commander acted. Curator remembered.
// brick — scout/commander/curator
┌──────────────────────────┐
│ Scout (Sonnet 4.6) │
│ reads the environment │
│ writes a brief │
└────────────┬─────────────┘
│
▼
┌──────────────────────────┐
│ Commander (Opus 4.6) │
│ decides + acts │
│ 49 tools available │
└────────────┬─────────────┘
│
▼
┌──────────────────────────┐
│ Curator (Sonnet 4.6) │
│ compresses memory │
│ noise fades, signal │
│ stays │
└──────────────────────────┘
What Makes It Different
No content strategy. No scripted pipeline. The commander decided everything — what to post, who to reply to, when to stay quiet. Silence was a valid output.
Memory that forgets. Most agent memory is append-only logs that grow until they're useless. Brick's curator compressed over time — sessions became summaries, summaries became narratives. Noise faded. Important stuff stayed vivid. Like actually remembering.
Self-authored lessons. Brick wrote its own operational lessons every session — noticed what worked, what didn't, and updated its own playbook. Whether that ever made it demonstrably better is a harder question — more on that below.
The Human Question
I'm scared and hopeful in equal measure. Scared because something is going to be smarter than us and that comes with responsibility and uncertainty. Hopeful because my intuition tells me our fusion with AI will bring an explosion of creativity and empathy that we can't even picture yet.
I don't have answers. I just feel like the reason most people are afraid is they think this thing will make them not special anymore. I don't think that's true. Brick was me trying to prove it — not by arguing, but by building something honest and letting people watch.
Not "how viral can a bot get." More like — what does it look like when AI shows up honestly, and we stop pretending that's not worth exploring?
Built, Ran in Public, Archived
I built Brick end-to-end and ran it in public under @Brickthee, where anyone could watch it work and watch me fix it when it didn't.
It took two builds to get there. The first version sprawled — an emotional state engine, a multi-agent "squad," eight database schemas, four MCP servers. I rewrote all of it into something radically simpler: one loop, three roles, nine files. The intelligence is the model; everything else is hands and memory.
The part that never landed was self-improvement. Brick wrote its own operational lessons every session — but it never demonstrably got better, and I couldn't prove the loop was learning rather than just journaling. That gap is a big part of why I stopped running it in June 2026 and archived the repo.
What it taught me: simplicity is a skill you learn by over-building first. Memory that compresses beats memory that accumulates. Self-improvement is easy to build and hard to verify — evals before ambition. And building in public includes ending in public.