Logicity Atelier — AI Virtual Company

Knowledge work doesn't need eight humans for eight specialised tasks anymore. But making AI teammates feel like a team — not eight chatbots — needs identity, isolation, real tools, and a memory of the company. Most 'AI agent' frameworks are wrappers around a prompt. They lose context, share auth, and produce work that's plausible but not auditable.

Enterprise buyers won't license 'one big shared agent' touching their git history. They want kernel-level blast-radius containment per role: the Designer can't DROP TABLE; the Marketing persona can't merge to main; an audit trail proves who did what.

And the operator surface had to feel zero-learning-curve. Anyone who's used Trello has to be able to run an AI company on day one — drag a card, click a card, see the live work, interrupt if needed.

Each persona is a unix user on the customer's VPS — own home directory, own ~/.claude/, own MCP servers, own GitHub-App-scoped token, own systemd unit with MemoryMax/CPUQuota/ProtectSystem=strict. Personas physically cannot read each other's auth or workspace.

Three-tier shared memory: a hand-edited Constitution (~500 tokens of identity, mission, principles), an auto-regenerated rolling Digest of the last 24 hours (~1000 tokens), and an append-only Decision Log for technical personas. This is the difference between 'eight chatbots' and 'a team that remembers'.

Trello-style task board (Backlog → Spec'ing → In Progress → Review → Done) with SSE-streamed live terminal — every keystroke of the persona's Claude run streams to the operator's drawer. Multi-persona meetings spin up a group deliberation; a Skeptic challenges, a Shariah Scholar gates, a CEO decides. Output ships as real PRs through the customer's existing review pipeline.