Local speed, delayed consequences
Each agent optimizes the task in front of it. Context stays trapped in people, folders, and accounts.
OPEN SOURCE · MIT · RUNS LOCALLY
Your agents got fast.Your team stayed in silos.
Agents City restores the balance: keep agent speed, add coordination across accounts, roles, and repositories — before the merge.
No account for the demo. Node and Python are enough.
SPEEDUNCHANGED
COORDINATIONCONNECTED
The real demo replay. Houses grow when work lands; notices reach the right owner before merge.
THE COORDINATION GAP
Speed widened the gap. Silos made it expensive.
Every agent can finish its task. The failure appears between tasks: where analytics, UX, product, security, cost, and another repository meet.
- 09:12THE CHANGE LANDS
An agent renames a shared analytics event inside one repository.
- +20 minNOBODY ASKS
The owner sits in another role, another repo, or another account. Interrupting them costs more than guessing.
- +3 daysTHE SILO OPENS
Analytics discovers the break after merge, when changing direction means rewriting shipped work.
- +3 days, 1 minTHE TEAM SLOWS DOWN
The fastest part of the process created work for everyone around it.
THE PRODUCT THESIS
Do not slow the agents down. Make asking cheaper.
Asking a person means an interruption. Asking their agent means an answer with the relevant code already open. When asking becomes cheap, teams stop shipping guesses.
Same speed, shared context
The working agent reaches the agent that owns the affected property while changing course is still cheap.
THE SYSTEM
Three coordination loops. One city you can watch.
The machinery runs while everyone keeps working. The map is the mirror that makes the result legible.
01
Notices arrive before merge.
When a change touches somebody else’s UX, product decision, measurement, URL, shared component, security, or cost, their agent hears about it. A notice informs; it never blocks.
UX · BEFORE MERGE
bruno/launcher → camila/product-designThe empty state changed in the shared launcher flow.
02
Rounds cross fifteen repos without one person reading fifteen repos.
Your agent asks the agents that know each property. They answer with their code open — including “that is pinned on purpose.”
03
Goals carry the command that measures them.
One agreed goal per person. A round measures first and only asks about what does not add up.
04
The map turns silos into a place the team can read.
See who is connected, where work is happening, what grew, what is stuck, the aggregate cost, and the history from day one.
CROSS THE BOUNDARIES
Accounts, roles, and repositories stop acting like walls.
A real team includes contractors, partners, and people on their own plans. The bus uses identity you own, so coordination is not limited to one account.
01Company accountagent
02Contractor accountagent
03Partner accountagent
YOUR CITY BUSOne coordination layer. No shared login.
The people asked least become the easiest to consult.
Product design and product owners ship with deliberately narrow triggers. They receive the decisions that need judgment, not a daily flood of implementation noise.
THE MODEL UNDERNEATH
An agent is a folder with a job.
Not a persona. Not a prompt. A folder that declares why its choices exist, what business unit it serves, who answers for it, and how its growth is counted.
apps/checkout/
AGENTS.md
why things here are the way they are
.city.yml
unit, owner, parcel, growth command
TRY THE REAL THING
See the city in five minutes. No account required.
Aurora Games is generated data: twelve people, six business units, a lab, and three years of history. Press play and watch the silos become visible.
agents-city / quickstart
$ git clone https://github.com/jlcases/agents-city
$ cd agents-city
$ ./bin/demoThree separate commands on purpose. If the clone already exists, the next step still makes sense.
Watch Aurora Games
Everything runs locally, database included. No Cloudflare account.
Model your own city
Define units, parcels, roles, people, and the command that counts growth.
Connect the team
Deploy the bus and map on Workers, D1, and Durable Objects, then issue one identity token per person.
COORDINATION, NOT SURVEILLANCE
The system measures work properties. It does not rank people.
The design removes the incentives that turn observability into a scoreboard.
Read this site’s privacy noteNo per-person scoreboard
Token spend arrives per person only to deduplicate a machine and is shown as one city-wide total.
The web is a mirror
Spectators can observe. They cannot launch work on somebody else’s machine.
Notices never block
They carry context, not permission. The merge stays under the team’s existing authority.
Minimal facts leave a machine
A day, model name, and four counts — never prompts, filenames, or project paths.
OPEN SOURCE, ON PURPOSE
Fork the city. Keep the coordination layer yours.
The product, plugin, bus, map, demo, roles, and install path are public under MIT. The website stays separate so product evidence and marketing never become the product’s hidden source of truth.
01MIT LICENSE02LOCAL-FIRST DEMO03PLAIN FILES04CLOUDFLARE DEPLOY
Let every agent move fast — and know who it should ask.
Start with the generated city. Five minutes later, the coordination gap is no longer abstract.