CURATED TIMELINE · EDITORIAL EDITION

From Browser Operators to Work Agents: The Manus Effect

Operator proved an agent could drive a browser. Manus made the broader contract legible: assign an outcome, let a cloud computer work, and receive a finished artifact. Follow that idea through ChatGPT agent, Manus 1.5, Claude Cowork, WorkBuddy, ChatGPT Work, and Sites.

Timeline overview

At a glance

Hover for the event briefScroll horizontally

2025.012026.07

01.23Operator 03.06Manus 07.17ChatGPT agent 10.16Manus Web App Builder 01.12Claude Cowork 05.28WorkBuddy 07.09ChatGPT Sites

8 nodes2025.01—2026.07 · Editorially curated · Updated 2026-07-15

Editorial thread

The timeline

Reviewed event briefs and original editorial context, ordered to show how the story changed over time.

01
2026-07-09Product preview
OpenAI launches ChatGPT Sites in public beta
ChatGPT Sites turns prompts or existing projects into hosted websites, lightweight apps, and games that users can refine, deploy, and share from ChatGPT.
Editorial context
The competitors finally met on the same end-to-end chain. Sites could start from a prompt or compatible local project, then generate, preview, revise, version, deploy, and share a website, app, or game with D1 data, R2 storage, authentication, and environment variables.

Both Sites and Manus 1.5 pursue nontechnical users who want a full-stack app from natural language. Manus emphasizes code export, custom domains, SEO, analytics, and Stripe; Sites emphasizes Work and Codex, local projects, ChatGPT identity, and Cloudflare-backed hosting. Similar category, different product boundaries — not evidence of plagiarism.
**The competitors finally met on the same end-to-end chain.** Sites could start from a prompt or compatible local project, then generate, preview, revise, version, deploy, and share a website, app, or game with D1 data, R2 storage, authentication, and environment variables. Both Sites and Manus 1.5 pursue nontechnical users who want a full-stack app from natural language. Manus emphasizes code export, custom domains, SEO, analytics, and Stripe; Sites emphasizes Work and Codex, local projects, ChatGPT identity, and Cloudflare-backed hosting. Similar category, different product boundaries — not evidence of plagiarism.
Read event
02
2026-07-07Product update
Claude Cowork expands to web and mobile
Claude Cowork is expanding to web and mobile, allowing delegated tasks to continue when users leave their computers and to resume across devices.
Editorial context
Cowork became a persistent work layer. Web and mobile access allowed a task to keep running away from the desktop and move across devices. Anthropic also said more than 90% of Cowork usage was outside software development, led by business operations and content creation.

That usage claim matters because it shows a coding-agent harness crossing into general knowledge work, not merely appearing in a launch demo.
**Cowork became a persistent work layer.** Web and mobile access allowed a task to keep running away from the desktop and move across devices. Anthropic also said more than 90% of Cowork usage was outside software development, led by business operations and content creation. That usage claim matters because it shows a coding-agent harness crossing into general knowledge work, not merely appearing in a launch demo.
Read event
03
2026-05-28Global launch
Tencent launches WorkBuddy office agent worldwide
Tencent Cloud took its WorkBuddy office agent global, dispatching tasks and sub-agents in parallel from a single instruction to deliver ready-to-use work output.
Editorial context
The office ecosystem became the platform. WorkBuddy was positioned as an agentic workspace rather than a smarter chat box, dispatching tasks and sub-agents while connecting models, enterprise data, and messaging clients.

Competition now included permissions, remote control, local files, and collaboration channels — the operational surface around the model, not only its benchmark score.
**The office ecosystem became the platform.** WorkBuddy was positioned as an agentic workspace rather than a smarter chat box, dispatching tasks and sub-agents while connecting models, enterprise data, and messaging clients. Competition now included permissions, remote control, local files, and collaboration channels — the operational surface around the model, not only its benchmark score.
Read event
04
2026-01-12Product preview
Anthropic launches the Claude Cowork research preview
Anthropic extends Claude Code's execution model to non-coding work, letting Claude plan tasks and produce professional deliverables inside authorized folders.
Editorial context
Cowork's direct lineage is Claude Code. Anthropic brought the coding agent's execution harness to files and knowledge work: authorize a folder, let Claude create and edit artifacts, split tasks, and deliver documents, spreadsheets, or presentations.

The ancestry is different from Manus, but the interface contract converged: describe the result, allow autonomous steps, then review the finished work. Convergence does not establish copying.
**Cowork's direct lineage is Claude Code.** Anthropic brought the coding agent's execution harness to files and knowledge work: authorize a folder, let Claude create and edit artifacts, split tasks, and deliver documents, spreadsheets, or presentations. The ancestry is different from Manus, but the interface contract converged: describe the result, allow autonomous steps, then review the finished work. Convergence does not establish copying.
Read event
05
2025-10-16Product launch
Manus 1.5 launches a full-stack Web App Builder
Manus 1.5 turns conversational site-building into a full-stack Web App Builder that deploys directly, covering backend, database, auth, and version management.
Editorial context
Manus turned website delivery into a product line. The 1.5 release moved beyond static pages to hosted applications with a backend, persistent database, authentication, embedded AI, notifications, visual editing, rollback, permissions, and custom domains.

This, not the abstract idea of a general agent, is the closest baseline for comparing ChatGPT Sites.
**Manus turned website delivery into a product line.** The 1.5 release moved beyond static pages to hosted applications with a backend, persistent database, authentication, embedded AI, notifications, visual editing, rollback, permissions, and custom domains. This, not the abstract idea of a general agent, is the closest baseline for comparing ChatGPT Sites.
Read event
06
2025-07-17Product launch
OpenAI merges Operator and deep research into ChatGPT agent
ChatGPT agent unifies web operation, deep research, code execution, and editable deliverables into a single agent.
Editorial context
The large platforms began joining separate capabilities into one task loop. ChatGPT agent combined Operator's browser control, deep research's synthesis, and code execution, with editable slides and spreadsheets as explicit outputs.

That is not proof that OpenAI simply followed Manus; Operator and deep research predated it. The shift is in the yardstick: after Manus, the market increasingly judged an agent by whether it could finish work end to end in its own environment.
**The large platforms began joining separate capabilities into one task loop.** ChatGPT agent combined Operator's browser control, deep research's synthesis, and code execution, with editable slides and spreadsheets as explicit outputs. That is not proof that OpenAI simply followed Manus; Operator and deep research predated it. The shift is in the yardstick: after Manus, the market increasingly judged an agent by whether it could finish work end to end in its own environment.
Read event
07
2025-03-06Product launch
Manus launches a cloud-based general AI agent
Manus asks users for an outcome, then plans, researches, codes, operates websites, and returns finished artifacts from a cloud virtual machine.
Editorial context
Manus changed the contract more than the underlying model. Users specified an outcome, while a cloud computer planned, searched, coded, operated sites, and returned reports, spreadsheets, or web pages.

The invitation scarcity and “first general agent” language were marketing. Calling it only a wrapper, however, misses the product innovation in orchestration, execution state, process visibility, and artifact delivery.
**Manus changed the contract more than the underlying model.** Users specified an outcome, while a cloud computer planned, searched, coded, operated sites, and returned reports, spreadsheets, or web pages. The invitation scarcity and “first general agent” language were marketing. Calling it only a wrapper, however, misses the product innovation in orchestration, execution state, process visibility, and artifact delivery.
Read event
08
2025-01-23Product preview
OpenAI previews Operator, an agent that controls a browser
Operator uses a visual computer-use model to click, type, and scroll through web tasks, handing control back for logins, payments, and sensitive actions.
Editorial context
Manus did not invent agents that operate interfaces. Six weeks earlier, Operator could inspect a browser, click, type, and scroll, with user takeover around credentials, payments, and sensitive decisions.

Operator's boundary was still the browser. It proved visual computer use, but it did not yet package research, code, files, and polished artifacts as one general work environment.
**Manus did not invent agents that operate interfaces.** Six weeks earlier, Operator could inspect a browser, click, type, and scroll, with user takeover around credentials, payments, and sensitive decisions. Operator's boundary was still the browser. It proved visual computer use, but it did not yet package research, code, files, and polished artifacts as one general work environment.
Read event