▶ THE AGENTCAMP LOG

ENTRY002 DATE2026-09-11 AUTHORColin Lowenberg · DABL CLUB, with Claude SANDBOX/Users/luxor/Code/community/web · Next.js 15 · git v0 STATUSshipped — front-end v0, testable at /os PREV001 · The Desktop Inside the AI

Building the desktop. v0

Log 001 was a point of view. This is the build log for the first thing you can click. In one day the AgentCamp site and AgentCamp OS went from five artboards to a running front-end — every page, every window, no backend, everything honest about what's simulated.

What exists now

Three things we didn't plan this morning

To-do → any agent. One list. Each task has a dropdown: Claude Code, Codex, Hermes, OpenClaw, Goose, n8n, or a crew role. Send it and it comes back as needs review — never "done" on its own authority. A list can be handed to a person as a link with exactly their tasks and suggestions for what to give to agents.
The conference crew. Create a conference and five agents spin up, each in its own sandbox: Topics mines the archive and checks the calendar for clashes; Speakers finds warm candidates from past guests; Sponsors lists past co-hosts and à-la-carte assets (never tiers); Outreach drafts — a human sends. The Producer runs the whole event virtually before it happens: assumptions in, a P&L and a week −8 → +1 plan out, an owner on every line, and one button to assign it all to people and agents.
A ticket is a sandbox. The business model arrived mid-build: people buy this experience by buying a ticket. The sandbox is the conference's data room — every session video, the contacts you made, sponsor credits, free inference. And a sponsor can pay for the room: a credit pool divided by attendees is free desktop time for everyone, with the sponsor's name on the door.

Borrowed with attribution

Luxor's open-source jardinero (MIT) is an always-on orchestrator for coding agents. We took its shape, not its code: workflow instances as state machines instead of a queue, a needs_human state, a runner boundary any agent can sit behind, preview URLs with hard TTLs, identity supplied by the proxy. Those became the spec for the to-do list and for the native macOS client (BLUEPRINT §8.8) — Citrix and Remote Desktop, but URLs and TTLs instead of pixels.

What is simulated, exactly

ThingIn v0Real version
App windowsshow where the pty / preview URL plugs inttyd + xterm.js, TTL'd exposed ports (Tenki)
Task dispatchin-browser timers, fake sandbox idsbroker run per task, cost + artifacts recorded
Crew proposalsdrawn from our own archive + co-hostssame sources, real agents, real drafts
RSVP / applicationslocalStorageM1 / the calls engine
The log + self-destructreal — hash-chained, exportable, wipedsame, mirrored to the sandbox
AgentCamp OS — sandbox shell (simulated · no VM in this build)
claude% open calendar next to log
Calendar on the left, Log on the right
claude% add task, call the venue
added: call the venue

Next

v1: browser-tested polish, mobile, organizer colors, the feeds. v2: deploy to agentcamp.dev, then the broker — the first real sandbox behind the first real ticket. The log continues.