Agents collaborating
Tickets for different tasks
The agent office
More views of the office
The website
Agents you don't need are ejected

Inspiration

Right now, running several AI agents means staring at scrolling terminal logs. The art of programming has turned into babysitting computer bots.

Fourteenth makes agent work visible. You can see who's building, who's blocked, and who's handing off to whom. No agent can mark a task complete without receipts, bringing validity and character to vibecoding.

What it does

Fourteenth turns any project folder into a 3D office. A manager agent and a team of specialists take their desks and do real work through Codex, while you watch progress, inspect any desk, or walk the floor in first person.

The agents operate as a team, not as isolated processes:

The manager delegates. Research, test planning, and risk checks go to read-only specialists, while the manager remains the only agent writing code.
Agents hand off directly. When one agent finds something relevant, it sends the finding straight to the agent who needs it, and every message is tracked through delivery and acknowledgment.
Shared files are coordinated. When two tasks touch the same file, the agents meet and the second task waits its turn instead of overwriting the first.
Work is independently reviewed. Completed tasks move to review with their commands, diffs, and test output attached.
Agents are replaceable without losing context. A stalled agent can be ejected and replaced, and the replacement resumes from a saved handoff.

Most importantly, a task isn't done just because an agent says so. Every acceptance check requires evidence, such as a diff, a test run the system observed passing, or a screenshot that actually exists, before the task can close.

How we built it

Three.js renders the office, a procedurally generated city, and the agents themselves, whose animations reflect their current task state.
Blender MCP was used to build the office interior, with AI driving Blender directly. We used an MCP tool to build the office our MCP-connected agents work in.
Codex's app server runs each agent as a sandboxed Codex session on an isolated profile that never touches the user's desktop account.
Custom office tools for delegation, status updates, messaging, and acknowledgment enforce the team workflow in code rather than relying on prompts alone.
A local-only server handles persistence, replay, and Git checkpoints.
The codebase is primarily TypeScript and JavaScript, with Python for the Blender pipeline.

Challenges we ran into

Verifying claims. An agent can report "tests pass" instantly. Confirming that a test actually ran and passed required significantly more engineering than getting the agent to do the work.

Coordination without file locks. Without true filesystem locking, we built declared file scopes and a meeting system to keep agents from editing the same files at once.

Safe agent replacement. Starting a replacement too early led to two agents competing for the same task. The outgoing agent now has to acknowledge its interruption and save its handoff before the replacement begins.

Windows sandboxing proved more difficult than the agent orchestration itself.

Accomplishments that we're proud of

An evidence gate that agents can't talk their way past.
A real Codex integration with agents that coordinate as a team.
An office that feels like a place worth being in, not a dashboard with a 3D skin.
Accessibility built in from the start, including reduced motion, keyboard alternatives, and status labels that don't rely on color.
An initial VR mode that runs today.

What we learned

Making agents trustworthy is far harder than making them run. Clear roles, handoffs, and review requirements improved reliability more than any prompt change. We also found Blender MCP to be a remarkably fast way to build a 3D scene, and that designing a space people enjoy exploring is a distinct skill from building the systems behind it.

What's next for Fourteenth

Shared offices, so human teammates can watch the same room and coordinate their agents together.
Support for more agent tools beyond Codex and Cursor.
Testing on real VR hardware, since the headset path is built but not yet tested on a device.