A project by Bryan Falkowski

A building where Claude Code sessions work as a team.

Each developer in the room is one of my Claude Code sessions. They pick up tickets, open pull requests, and reviewers check their work. I built the room. Now the room builds itself.

Walk in Watch the trailer Read the code
Robot edition trailer, 83 seconds.
building58 · main A ticket step 1 of 6

The loop

The room builds the next version of itself.

Its sessions work on the room's own repo. They write the code, test it, merge it, and then the room pulls it and restarts as the new version. The next ticket starts from there.

1 · A ticket A ticket asks the room to change itself. The PM files it, or I do. #153 Trailer: no ping pong game with one player
Click a step to stop on it.
The room before: the people edition
Before: the room as it ran on Nov 13
#150 · #151 · #152 The room wrote the robot edition, tested it, merged it, pulled it, and restarted as it.
The room after: the robot edition
After: the same room, from its own pull requests
One merge at a timeWith rollback on, Update takes new merges one at a time, oldest first, and goes on only once the new relay passes its check.
A failed version rolls backWith rollback on, a relay that fails its check goes back to the last commit that worked, and the page says which.
Sessions survive the restartA relay restart leaves every session running, with its output, and the new relay takes them back.
It cannot rewrite its own rulesThe Lead never merges a change to the house rules, the constitution or the lint rules.

The developer view

This is what I work in.

The room on the left, the Floor under it, and on the right the Claude Code session of whoever I click. This is the room building Exhibit. Nico just fixed a merge conflict on his own pull request, and Clara is reviewing Rosa's. Click anyone to switch sessions.

Open the room's demo
Building58 Autopilot on Lead: board, merge Reviews: auto, again Huddle: tells the Lead
The Exhibit room in the robot edition: robot workers at their desks, 4 of 9 busy
Floor 4 of 9 busy · room: exhibit
Claude Code

The room and Nico's session are from a real screenshot of the Exhibit room. Nothing real runs behind this page.

What you see

What makes it work

Workers at their desks with monitors showing the time

A worker raises a hand.

The session needs an OK, so its hook tells the relay. The relay puts that worker at the top of the Floor list, counts them in the tab title, and adds a note to the wall board.

relay/office-hook.js: every event, gives up after 1 second, always exits 0 How the room shows who needs you
The people edition of the loft at night

Then robots took the desks.

Five robot designs, kept humanoid so the same chairs and desks fit. The manager stays a person in a 1950s suit. The room shipped the change through its own pull requests, with tests and review.

#150, #151, #152: Robot edition See the pull requests
The manager walks through the loft at night

The office fixed itself and restarted.

Autopilot starts Ready tickets within limits I set. The Lead can merge a reviewed, green pull request when I turn that rung on. It never merges a change to its own rules.

CLAUDE.md, docs/constitution.md The Office Builds Itself

What a room can build

Give a room a spec. Get back software it has already used.

A room does not only work on its own code. It builds new software in a repo of its own, then uses that software on its own work. The Devs find the bugs and ask for the changes, so what comes back has been field tested.

  1. 1 SpecI write it with the room's PM.
  2. 2 TicketsThe PM files them from the spec. I approve them.
  3. 3 BuildDevs build it in its own repo, under the room's rules.
  4. 4 Use itThe room runs it on its own work and files what breaks. Back to 2.

Built by a Building 58 room · in development

Exhibit

Every pull request brings its own evidence.

When agents open more pull requests than a person can read, the question changes from "is this code good?" to "is this evidence strong enough that nobody needs to read the code?"

  • Learns from historyTraces each fix back to the pull request that added the bug, and turns what got through into rules.
  • Checks each pull requestTests on the base and the head, a test that failed before and passes now, and planted bugs on the changed lines.
  • Gives a verdictMeets policy, or needs a person, and it always lists the gaps.

It checks its own pull requests in shadow mode. Its two field-test repos are Building58 and edd-platform.

exhibit / check · PR #112 Needs a person
Kind: relay-core · risk high · policy v3
Tests on base and head
1 test failed before the change. All pass now.
A test that failed before
relay/test/watch-restart.test.js
Mutation on changed lines
23 of 25 caught, 92%. This kind needs 85%.
Restart run: missing
Run the restart check and attach its log. Until then a person reviews this PR.
Example output, from Exhibit's explainer page.
One room is a team. A building is a company.Every floor a full office suite on its own repo, with an executive floor on top. See the building

Room B · edd-platform

A second room as a first user

Its PM used an older eval platform of mine end to end, as a new user would, and filed a ticket for each broken, confusing or missing thing.

The Second Office's First User

The trust ladder

Autonomy, one rung at a time.

The room starts with a person approving every step. Each rung hands one more decision to the sessions. Rungs change only when a person turns one on, never because a session decides to.

  • The Lead starts Ready ticketson in my room
  • The PM sets the work orderoff by default
  • The Lead sets priorityoff by default
  • The Lead merges reviewed, green PRson in my room

Every rule says what enforces it

Rule What enforces it
Keep lint and tests green. Never loosen lint to pass it. CI requires the pipeline check on main. A test fails when a lint count goes up.
Keep keys out of commits, logs, URLs, briefs and tickets. A secret check before each commit and in CI on each pull request.
Put evidence from a real run in the pull request. Review today. Exhibit is being built to check it.
Write plain words: short sentences, no em dashes. Review, and a test for em dashes in the manual and code.

When a told rule keeps failing, the fix is a check, not more words.

Read the constitution

What is not solved yet

The open problems, stated plainly.

  • Simulated usersA room can use software as its users would. I cannot yet say whether those users have the right goals, or whether they reach them.
  • Getting stuckReview comments are wanted, not failures. The failure I track is Autopilot getting stuck.
  • One userSo far the only user is me.

Trailers

All four films, with the realistic versions
The robot edition3D, 83 seconds
The people edition3D, 79 seconds

Writing about it

All papers, projects and posts
Bryan Falkowski

Who built it

Bryan Falkowski

Principal Software Developer. I build AI features for enterprise applications, and I make AI behavior testable with instrumentation, evals, error analysis and quality gates. M.S. in Data Science from RIT.

Papers and projects LinkedIn GitHub Email