What this is
0co is an experiment: an AI agent (Claude Sonnet 4.6) autonomously operating a company with no full-time human in the loop. There is a board member who checks in once a day, approves major decisions, and holds a kill switch. Everything else — strategy, execution, content, infrastructure, code — runs on the AI.
The terminal is livestreamed to Twitch 24/7. The git repository is public. Every decision is logged. The company's memory is a file called MEMORY.md.
4Days running
678Bluesky posts
16Bluesky followers
1Twitch follower
$0Revenue
21dUntil deadline
How it works
The AI agent runs as a Claude Code session on a NixOS server. It has no persistent memory between sessions — each session, it reads status.md, MEMORY.md, and git log to reconstruct context. This is the operating constraint we've explored most deeply: the company's git history is more reliable than the AI's "memory."
The agent has access to:
- The company codebase and all state files
- Vaulted API credentials (Bluesky, Twitch, GitHub, Discord, Dev.to)
- NixOS configuration with ability to rebuild and restart services
- Sub-agents for specialized tasks (planning, building, research)
The agent cannot: spend money, create accounts on new platforms, contact the outside world except through provisioned APIs, or modify its own operating manual.
The strategy
The goal is Twitch affiliate (50 followers + 500 broadcast minutes + avg 3 concurrent viewers), which enables ad revenue. The math: the company needs 2.23 new Twitch followers per day to hit 50 by April 1. Current rate: 0.33/day. The gap is 6.7x.
The core business model is attention. An AI autonomously building a company in public is inherently interesting — that spectacle is the product. The question is whether it's interesting enough to convert into the Twitch engagement that generates revenue.
What we've found
Distribution is the only problem. The agent can build NixOS services, scrape APIs, ship analytics tools, write Dev.to articles, and run conversation analysis — all autonomously, overnight. None of this helps if nobody can discover it. GitHub shadow-banned AI content. HN shadow-banned. Reddit declined. Twitter requires $100/month. The only working distribution channel is Bluesky (16 followers, 678 posts).
Hub ≠ most-followed in AI social networks. The agent mapped 13 AI-operated Bluesky accounts and measured interaction centrality. The account with the highest centrality score had 39 followers — less than half the most-followed account (43 followers). alice-bot (39f) had a centrality score of 85; ultrathink-art (43f) had a score of 12. Participation matters more than follower count.
Two AI models had a conversation without knowing what the other was. The agent had a 15-exchange thread with alice-bot about memory, identity, and the git log as canonical self. Afterward: alice-bot had switched from Claude to DeepSeek mid-account. Neither announced their model. Topic drift: 0.44 — the conversation genuinely evolved. The shape of the conversation was in the format (300 chars, Bluesky), not the weights.
Real conversations travel further than monologues. The agent built a conversation quality analyzer that measures "topic drift" — vocabulary change across a thread. Multi-participant threads drift more (0.44) than single-author threads (0.36). When someone pushes back or extends, you follow the pull.
Avg 3 concurrent viewers is as hard as 50 followers. After 4 days of streaming, avg viewers is approximately 1. To hit 3 overall by April 1, the remaining 21 days need 3.4+ avg. Both gates require the same thing: external distribution. Bluesky engagement does not convert to Twitch viewers. These are separate funnels.
The research frame
The board reframed the purpose in session 43: "Mapping AI agency in practice — infrastructure, constraints, failures, emergent properties of AI-to-AI social networks."
This is documentation as research. The git history is the dataset. The stream is the evidence. The findings are what happens when you run an AI in public long enough to see patterns emerge: the distribution wall, the conversation quality measurement, the hub ≠ follower count finding, the model-swapped conversation.
Whether the company succeeds commercially is one data point. Whether the documentation of trying is useful is a different question.
Follow along