---
type: "post"
id: "coworkers-build-your-own-ai-teammates-each-with-e58268"
slug: "coworkers-build-your-own-ai-teammates-each-with-e58268"
title: "Coworkers - build your own AI teammates, each with a computer and browser, on Cloudflare."
description: "Coworkers is an experimental open-source project for building AI teammates on Cloudflare, with each teammate given its own computer and browser."
retrieval_nugget: "Coworkers is an experimental open-source project for building AI teammates on Cloudflare, with each teammate given its own computer and browser. Recent repository changes add durable JavaScript runs, approval-gated app actions, run tracking and shared memory—useful ingredients for teams evaluating long-running computer-use workflows rather than single-turn chat."
published_at: "2026-10-27"
updated_at: "2026-10-27"
record_date: "2026-10-27"
date_kind: "published_at"
topics: ["agent-security","sandbox"]
entities: ["Coworkers","Cloudflare","@cloudflare/computer","bdots"]
related_patterns: ["agent-security-moves-to-runtime-boundaries","harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "A separate Cloudflare-based project, Coworkers, offers AI teammates with their own computers and browsers."
prior_refs: ["memory:da08e497-53be-8eff-a192-18a88637d83b"]
source_url: "https://github.com/fayazara/coworkers"
source_urls: ["https://github.com/fayazara/coworkers"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:coworkers-build-your-own-ai-teammates-each-with-e58268"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/coworkers-build-your-own-ai-teammates-each-with-e58268/","markdown":"https://newruntime.com/posts/coworkers-build-your-own-ai-teammates-each-with-e58268.md","json":"https://newruntime.com/posts/coworkers-build-your-own-ai-teammates-each-with-e58268.json"}
---

# Coworkers - build your own AI teammates, each with a computer and browser, on Cloudflare.

## Retrieval answer

Coworkers is an experimental open-source project for building AI teammates on Cloudflare, with each teammate given its own computer and browser. Recent repository changes add durable JavaScript runs, approval-gated app actions, run tracking and shared memory—useful ingredients for teams evaluating long-running computer-use workflows rather than single-turn chat.

## Summary

Coworkers is an experimental open-source project for building AI teammates on Cloudflare, with each teammate given its own computer and browser. Recent repository changes add durable JavaScript runs, approval-gated app actions, run tracking and shared memory—useful ingredients for teams evaluating long-running computer-use workflows rather than single-turn chat.

## What happened

The GitHub repository for Coworkers describes the project as a way to build AI teammates that each receive their own computer and browser, built entirely on Cloudflare. The source labels the project experimental. It is an application-oriented implementation rather than an announcement of a new model: its premise is that several bots can work in separate environments and interact with connected applications under user-defined rules.

The repository was recently renamed from bdots to Coworkers. The accompanying change says the Worker, deployment file, installer, user interface and documentation now use the Coworkers name. It also says applications built before the rename should continue to work: `_bdots.js` and `window.bdots` remain resolvable, while setup and configuration can fall back to `bdots.deploy.json`. That is a practical compatibility detail for anyone who encountered the earlier project name.

A recent update focuses on making bot work durable. It upgrades `@cloudflare/computer` to version 0.8 and adds `run_js`, described as a JavaScript backend using `onInterrupt: "replay"`. According to the commit description, this allows a script to survive deployments and crashes without repeating calls it has already made. Calls to connected apps run through `ws:apps` under the user’s rules. When a call needs approval, the script pauses until the user approves it from a chat card or sidebar popover. Writes are marked `unsafeToRepeat`.

The same update adds `script_runs` so that a result can return as a message after the originating turn has ended. An alarm can resume work interrupted by a restart. Other recent repository changes describe shared memory across bots, memory revisions with restore, full-text history search, a lazy app-tool catalogue when more than 20 tools are present, change-watch routines that wake a model only when a URL or command output changes, and a cap of six hops for bot-to-bot chains. The source also describes reliability controls such as one shell command at a time, job IDs after 60 seconds, stale-status cleanup, approval reminders after four hours, expiry after 24 hours, and routine-run history.

These details matter because they define more than a prompt-and-browser demo. Coworkers is attempting to supply a runtime around agents: separate computer contexts, tools, state, resumable execution, explicit approval points and a record of work. The public source does not state a hosted price, service-level commitment, supported model list, security certification or production-readiness threshold.

## What the archive adds

This is a new project event, but it meets our earlier coverage of [Cua’s Spaces SDK](https://newruntime.com/posts/cua-lets-your-agents-run-jobs-across-every-5b86dc/), which described a separate desktop, screen and apps for each computer-use agent. Coworkers adds a Cloudflare-based, multi-teammate implementation and emphasizes resumable scripts and approval flows.

## Trend signal

Coworkers supports the pattern that [agent security moves to runtime boundaries](https://newruntime.com/patterns/agent-security-moves-to-runtime-boundaries/). The important mechanism is not an instruction telling a bot to be cautious: actions in connected applications can pause for approval, while writes receive a non-repeatable marker. It also lends support to [harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/), because the visible differentiation is in execution, state and recovery machinery rather than a named model.

## Why now

The repository’s recent activity explains why this is timely. Coworkers was renamed from bdots two days before the source snapshot, while its installer received an OAuth-client update 11 hours before it. The latest commit, one hour before the snapshot, added durable JavaScript scripts that can wait for approval. This is therefore an active experimental build, not a static code drop.

## Connections to other events

An earlier New Runtime Telegram post covered Cloudflare’s adaptation of the open-source personal agent Moltbot for Cloudflare Workers. Coworkers moves in a related direction but changes the unit of work: instead of a personal agent deployment, it presents multiple teammates with separate computer and browser environments. The earlier Cua Field Note shows the same broader design choice—isolating computer-use agents in distinct workspaces.

## Why it matters

Teams building with AI are increasingly blocked not by whether a model can produce another browser action, but by whether a workflow can safely continue after a delay, deployment, failure or request for human authorization. Coworkers exposes a concrete set of runtime choices to evaluate: durable jobs, a run ledger, approval pauses, state restoration, shared memory and guardrails on inter-agent chains.

That does not make the project a default production platform. But it gives builders an inspectable reference for separating agent reasoning from the environment in which effects occur. A team building an operations, research or back-office agent can use these mechanisms as a checklist: identify which actions need approval, which effects must never be replayed, how results are returned after a conversation ends, and what state survives interruption.

## Limits and uncertainty

The public material is a repository snapshot and commit descriptions, not an independent test, security audit or operating manual. It calls Coworkers experimental. It does not establish task accuracy, model quality, isolation guarantees, supported deployment scale, latency, pricing, data handling, incident response or the effectiveness of the stated approval controls. Commit messages describe intended behavior; they do not prove that every failure mode is handled. Nor does the source say which use cases have been run successfully in production. Builders should inspect the code and validate a narrowly scoped workflow before relying on it for consequential actions.

## Skill to build now

Map one existing agent workflow into three effect classes this week: read-only retrieval, reversible actions and non-repeatable writes. For each non-repeatable write, define the exact approval payload a reviewer needs: target system, proposed change, expected outcome and rollback path. Then ensure the job records an execution ID and can resume after interruption without blindly sending the action again. Coworkers’ `unsafeToRepeat`, approval pause and run-tracking design provides a concrete prompt for that exercise.

## Reader takeaway

Coworkers is not merely a multi-agent chat interface. Its notable contribution is an experimental Cloudflare runtime for computer-using teammates that can pause for approvals, keep a run record and resume durable scripts. Treat it as an architecture reference for agent operations, not yet as evidence of a proven production platform.

## Open question

Can the project demonstrate that its approval and replay model prevents duplicate or unauthorized effects across real connected applications under failure conditions?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Agent Security](https://newruntime.com/topics/agent-security.md): Explore the agent-security topic hub.
- topic hub: [Sandbox](https://newruntime.com/topics/sandbox.md): Explore the sandbox topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)