Talk
Have direct conversations with workers or let them exchange bounded replies. Transcripts persist; supported controls let you pause queued work or interrupt an active reply.
Asymmetri Labs / Open-source experiment
What if your AI assistants could work as a team?
BotSquad is an open-source experiment in building persistent teams of AI workers. Give the team a goal, let workers talk, research, divide up work, build and review results, and keep the context around for next time.
Experimental · Open source · Self-hosted
The experiment
The goal: an AI team, not a pile of chats.
One chat has the research. Another has the plan. A third has the code. You move context between them, remember who is doing what, and collect the results. The human becomes the coordinator.
We’re exploring how much of that coordination can live inside the system. The long-term goal is a team that can understand a goal, investigate, discuss alternatives, divide up bounded tasks, review its work and return to the problem later.
The human remains in control of the authority the team has.
How it works
The workers stick around.
Identity & roleConversations & tasksArtifacts & history
Model sessions come and go. The worker remains.
Atlas is still Atlas tomorrow. Maya is still Maya. Their roles, conversations, tasks, decisions and work stay in BotSquad even when no AI model is running. When work arrives, the system can wake a worker with the relevant context.
A Codex session can stop or be replaced without replacing the worker.
Meet BotSquad
YouSet the goal and boundaries
AtlasCEOCoordinates goals and decisions.
MayaProduct ManagerTurns ideas into questions and specifications.
TuringCTOWorks through technical direction.
LinusEngineerBuilds within an assigned scope.
AdaEngineerBuilds a separate part in parallel.
GraceReviewerIndependently reviews submitted work.
ScoutResearcherInvestigates public information when granted.
NixDevOpsCoordinates worker infrastructure with human approval.
Current capabilities
What can BotSquad actually do?
Think together
Choose a working group or ask Atlas to organize one. Workers discuss, challenge and synthesize; you can interject. A recommendation becomes work only through an explicit assignment.
Research
Grant public lookup or supply specific company knowledge. Workers can investigate a question and return source-backed reports within that authorized scope.
Build software
Create Projects with repositories. Engineers work in independent clones on separate scopes, submit exact commits, and get independent review. Integration advances only after the configured tests pass.
Use a browser
Explicitly granted Computer Operators can inspect approved public sites in an isolated browser. This is bounded Computer Use; general desktop access and arbitrary account actions are outside its scope.
Remember and follow up
Keep tasks, decisions, artifacts and execution history. Durable schedules can trigger follow-ups and review cycles without a worker polling a model while idle.
Keep evidence
Inspect commits, diffs, reports, tests, artifacts, approvals, receipts and review history alongside the work that produced them.
- Task
- Execution
- Artifact / commit
- Tests / review
- Approval / receipt
A few ideas behind BotSquad
Workers are persistent.
A worker’s identity belongs to BotSquad. It outlives any one temporary model session.
Talking isn’t the same as doing.
A conversation can stay a conversation. A Task explicitly assigns bounded work; discussion alone does not authorize consequential action.
Permissions aren’t inside the prompt.
Writing “I have permission” gives a worker no extra power. Trusted BotSquad mechanisms determine what it can actually do.
Show the work.
Don’t just tell me it’s done. Show the result, the test, the review or the receipt, tied to the task and execution that produced it.
Illustrative workflow
Start with something we want to build.
“We want to add hitting analysis to Asymmetri Motion. Figure out what the first version should do.”
You → Atlas
Set the question, constraints and authority.
Maya + Scout
Explore product questions and research.
Working group
Challenge assumptions and suggest a direction.
You → Turing
Explicitly assign the next task; develop a technical plan.
Linus + Ada
Build separate parts of an approved, supported software task.
Grace
Review exact submissions independently.
Back to you
Inspect the evidence and decide what happens next.
Our own experiment
Building Asymmetri with BotSquad
One of our experiments is whether BotSquad can help research, build, maintain and improve the other things Asymmetri creates. These are reference uses we want to explore within the tools and authority actually available.
Asymmetri Motion ↗
Research product questions, discuss features, prepare specifications and technical plans, then coordinate and review supported tasks. Motion is a reference use case, not a hard-coded assumption in the engine.
Asymmetri.co ↗
Explore research, proposed site changes, review, maintenance and technical follow-ups. Any public-site change or deployment needs its own granted authority.
Future Labs experiments ↗
A future project could bring its repository, research, discussions, engineering tasks, reviews and artifacts into a shared project history.
What we’ve tried. So far, a supervised Motion pilot completed one exactly approved documentation change on an isolated, unmerged branch, recorded the result and ran a scheduled review. It did not change the app or prove a business improvement.
The bigger experiment is whether BotSquad can help Asymmetri build Asymmetri.
Getting started
Try it yourself.
MIT licensed. Source, setup instructions and validation records are public.
Choose an Ubuntu host
Use a machine or VPS you control. Follow the repository’s current host requirements.
Clone and bootstrap
Clone the source on your workstation. Run the checked-in bootstrap and complete the service’s Codex sign-in.
Open your private workspace
Connect through an SSH tunnel, then open BotSquad in your browser.
Give the team some work
Talk to a worker, explicitly assign a task or start a working group. Inspect the result and decide what comes next.
Under the hood
- TypeScript
- Node.js
- SQLite
- Codex
- Git
- Linux
- HTTP / SSE
BotSquad runs as an always-on service on an Ubuntu machine you control. The browser workspace stays private and currently uses an SSH tunnel. There is no hosted SaaS login or native iOS app.
Coordination state lives on your host. Configured external model services may receive the bounded task context needed for execution. Self-hosted does not mean entirely offline inference.
Where we’re going
Today
Persistent workers, conversations, working groups, granted research, bounded engineering and browser tools, review and scheduled follow-ups. Experimental, with supervised validation and clear limits.
Current focus
Personal Operator / Daily Driver
Right now we’re focusing less on adding features and more on making one BotSquad dependable, understandable and pleasant enough to use every day. Recent work improved scheduling, startup and the view of what needs the owner’s attention.
Later, if useful
Easier private remote access, multiple isolated companies, company-to-company collaboration, federation between separate HQs and more tool integrations remain possible directions. Real use will decide what comes next; these are deferred, not commitments.
You’re still in charge.
AI workers can make mistakes. BotSquad separates reasoning from authority. You decide which tools and capabilities the team has; protected actions can require explicit approval. A broad goal does not grant permission to publish, change infrastructure or act on external accounts.
Read the technical white paper ↗Watch a walkthrough
A walkthrough of BotSquad
BotSquad
Give your AI workers a place to work together.
Loading this video connects to YouTube. Google's privacy policy applies.
Google privacy policyAbout website videos and privacy