GUIDED DEMOAurora Games
How do we restore vanished saves without making people lose progress twice?
Every person already works with their agents. The company still needs them to work together.
Give each person one or more local cities with an accountable chair, specialist agents, their own knowledge, and explicit roads to other cities. The Apache 2.0 core works with code, documents, and operational work across any domain.
$npm install -g agents-city@beta$agents-city- Apache 2.0 core
- 9 built-in domains
- No telemetry or licence server
THE DOMAIN CHANGES
3 demos · 0 model accountsThe accountability protocol does not.
- 01isolated positions
- 02synthesis
- 03decision
- 04failed verification
- 05replan + pass
THE DEMO RUNS THE WHOLE STATE MACHINEThe first decision fails. The city finds the gap, replans, and closes only after verification.
01 · ISOLATED POSITIONS02 · CONFLICT SYNTHESISED03 · ATTRIBUTED DECISION04 · VERIFY AND REPLAN
The company did not lose productivity. It lost coordination.
One person with good agents can finish work before a cross-functional meeting even starts. So people stop asking each other. The output accelerates, but context, objections, and decisions stay trapped inside separate human-and-agent silos.
Each person and their agents move fast alone.
- Software
Implementation finishes before Product sees the choice.
- Healthcare
Clinical operations and patient safety inspect different evidence.
- Legal
The deadline, precedent, and client decision live in separate workspaces.
- Operations
A material dependency appears after the plan is already moving.
Their cities coordinate before humans need a meeting.
- 01Ask once
A chair turns the business question into a precise brief.
- 02Consult the right experts
Each one answers from their own files, tools, and context.
- 03Resolve the conflict
The chair compares evidence instead of counting votes.
- 04Return one outcome
Decision, owner, dissent, and independent verification are recorded.
In companies I see productive people move much faster with their agents, but their agents still work in isolation. Engineering can finish before Product is even asked. I built Agents City so each person’s city can consult the right colleagues’ cities before the work is already done.
The developer’s agent asks Product before it builds.
This is one real example, not the product boundary. The developer does not need to stop working, find a calendar slot, or paste half a conversation into Slack. Their city recognises a product decision and asks the right person’s city—even while that person is busy elsewhere.
QUESTION BEFORE IMPLEMENTATION
“I can move the trial limit into Checkout and remove one settings screen. It is simpler technically. Does this change the product promise?”
Evidence: diff summary · flow-preview.company.test/482EXPLICIT ROAD · TEXT ONLY→
CONFIGURED BY THAT PERSON
- Decision criteria
- Product knowledge
- Authorised data
- Existing skills and tools
PRODUCT CONTEXT
“Do not remove the settings screen. Existing administrators use it after purchase. Keep the implementation smaller and preserve that control.”
BACK IN ENGINEERING
The decision and rationale return to the developer’s city. Its agents work against the agreed product constraint, while both owners can review the exchange later in their own local session history.
✓Give every person several local cities. Define the agents inside each one.
A city is a local, accountable team—not another group chat or a list of repositories. Current main lets you add agents one at a time. Each one may work across several repos, worktrees, or document folders, use its own runtime, and start from the skills its owner has deliberately chosen. A city may also begin with only its chair.
One person can keep several boundaries.
Product, client work, research, or any other domain can remain a separate city with its own identity, goal, knowledge, agents, and roads.
The agent is the unit.
Name the specialist, choose whether its work is code, knowledge, or coordination, then mount everything it needs. One repo is only one possible source.
Each specialist keeps the right context and runtime.
An agent may combine several repos and document folders, then run on Claude, Codex, OpenCode, Kimi, or an explicit terminal fallback.
Skills are an owner decision.
Agents City recognises skills already present in mounted work. Current main can also install a folder or zip that the owner explicitly selects into a compatible agent home; nothing happens automatically.
When another city belongs to a colleague, client, or supplier, only the two chairs cross an explicit road. Agents, files, credentials, and execution stay on their owners’ computers.
In committee mode, an agent city works like a real human team.
This is the current mode for making a decision: not everyone talks to everyone or decides at once. Each specialist prepares a position independently; the chair brings the relevant views together, takes responsibility for the decision, and asks another agent to verify it.
- 01
Ask one precise question
State the objective, constraints, authority, definition of done, and who may contribute.
- 02
Choose relevant experts
Invite only the people or agents whose evidence can materially change the decision.
- 03
Collect positions in isolation
Nobody sees the first answer until every invited specialist has responded or abstained.
- 04
Synthesize the conflict
Compare evidence, disagreements, risks, and unknowns. Do not vote and do not average opinions.
- 05
Name the decision and owner
Record the outcome, rationale, executor, dissent, and conditions that would reopen it.
- 06
Verify independently
A different agent checks the result. Failure triggers a replan; only a pass closes the work.
Start with a city prepared for your work.
Choose software, healthcare, legal, finance, marketing, sales, research, or operations. Agents City suggests the initial specialists and checks; you can change everything or create your own domain.
Software
Coordinate product, engineering, data, design, and operations before a change ships.
Healthcare
Bring clinical direction, patient safety, operations, data, and compliance into one bounded decision.
Legal
Work a matter across deadlines, precedent, client context, compliance, and knowledge.
Finance
Coordinate close, controls, analysis, compliance, and operational cycles.
Marketing
Align brand, content, acquisition, lifecycle, design, and measurement.
Sales
Connect account evidence, revenue operations, enablement, and customer outcomes.
Research
Separate methods, evidence, ethics, knowledge, and verification before synthesis.
Operations
Resolve dependencies across programmes, processes, quality, knowledge, and delivery.
Your domain
Build a city around your own team, vocabulary, and way of deciding.
Your cities stay on their computers. You decide what network they form.
Local cities need no server. When work must cross machines, deploy the relay included in the repository to your own Cloudflare account. Every owner keeps their city, and every Road is explicitly allowed between two chairs.
OPERATED BY YOUR COMPANY OR COMMUNITYCloudflare WorkerWORKER INCLUDED IN COMMUNITY
Your own company
Connect Product, Finance, Legal, Operations, or any other city without merging their files, credentials, or goals.
Your clients and suppliers
Exchange questions and evidence across a declared boundary. Each organisation keeps its agents and may close the Road from its side.
CPS communities
Create a Complex Problem Solving community: developers and other specialists keep private cities and contribute only the answer or evidence they choose to share.
You do not share your knowledge base. You share what your city can contribute.
- 01Bounded question
The Road delivers one question to the chair of the other city.
- 02Local consultation
Its agents work with documents, data, and tools that never leave that machine.
- 03Chosen response
Text returns with the conclusion or evidence that city elects to contribute.
- seat ↔ seat only
- invitations contain no token
- 72 h maximum queue
- no message logging by default
Agents City does not operate this bus or receive its traffic. The guide uses the Worker, KV, and Durable Object already shipped in the Apache 2.0 repository.
Self-host the relay on CloudflareYour cities stay on your computers. You decide who they can talk to.
Agents City stores each city locally and sends no product telemetry to us. To coordinate with a colleague or client, both cities connect out to infrastructure your organisation runs. No one opens their computer to incoming access.
See what stays localYour city remains yours
Its configuration, state, and local knowledge remain on the owner’s machine. Agents City has no product account or cloud where we collect them.
Connect without opening a computer
Each side starts its own outbound connection to your relay. An invitation gives no one access to files, credentials, or execution.
Choose exactly who may talk
A Road connects two named chairs. It does not create an open network between every agent or every city.
Operate the shared path
Your organisation runs the relay, controls access, and can remove any Road. Message logging stays off unless you explicitly enable it.
The open core is the product. Enterprise starts with your problem.
The complete published core is Community Edition. When a real company constraint remains, Arkatai designs and builds the smallest custom extension around it.
Community Edition
Apache 2.0Use it, modify it, self-host it, and build on top of it. There is no product account, licence server, or telemetry.
- Local cities, Hall, agents, workspaces, and mounts
- Nine domain packs with editable roles and knowledge
- Committees with isolated positions, verification, and replan
- Local roads plus an optional relay you self-host
Enterprise, built to fit
Custom by ArkataiNo fixed bundle. We start from a real company problem and build the missing capability around the open core.
- Start from a real workflow, constraint, or governance problem
- Design the smallest extension that solves it
- Integrate with the systems and policies already in place
- Agree scope, evidence, ownership, and handover before building
The full product has a real manual.
Use the site to understand why Agents City exists. Use the docs when you need the exact command, file, boundary, expected result, or failure mode.
Build your first agent city.
Run the guided demo on your computer. Then create a local city with a real objective, define its agents, and mount the repositories, worktrees, or document folders each one needs.