# 1.5 How today's tools realize it (/ai-worker-paradigm/from-chatbots-to-ai-workers/both-ai-vendors) --- type: Document title: "1.5 How today's tools realize it" description: "How Anthropic, OpenAI and open-source tools realize the change today, the differences that change a decision, and which tool to use for which job." status: stable order: 101.5 ksor: owner: team:panaversity audience: [ public ] approval: by: process:panaversity at: 2026-11-05T06:58:08Z chapter: "01" part: I expert_status: required concepts: [ "1.5" ] last_verified: 2026-10-22 review_by: 2026-11-22 sources: - id: openai-moving-desktop title: "Moving to the new ChatGPT desktop app (OpenAI Help Center, verified 2026-11-04)" resource: https://help.openai.com/en/articles/20001276-moving-to-the-new-chatgpt-desktop-app - id: anthropic-one-claude-post title: "Claude Cowork and chat are now one Claude (Anthropic, 16 September 2026, verified 2026-11-03)" resource: https://claude.com/blog/cowork-is-now-claude - id: anthropic-one-claude-plans title: "Claude Cowork and chat are one Claude, where tasks run (Claude Help Center, updated 30 October 2026, verified 2026-10-31)" resource: https://support.claude.com/en/articles/16761823-claude-cowork-and-chat-are-one-claude - id: anthropic-one-claude-nov title: "Claude Cowork and chat are one Claude, on defaults and moving over (Claude Help Center, verified 2026-11-03)" resource: https://support.claude.com/en/articles/16761823-claude-cowork-and-chat-are-one-claude - id: anthropic-cowork-surfaces title: "Use Claude Cowork on web, desktop, and mobile (Claude Help Center, verified 2026-11-03)" resource: https://support.claude.com/en/articles/15520349-use-claude-cowork-on-web-desktop-and-mobile - id: anthropic-desktop-linux title: "Claude Desktop on Linux (beta) (Claude Code documentation, verified 2026-10-28)" resource: https://code.claude.com/docs/en/desktop-linux - id: anthropic-cowork-architecture title: "Claude Cowork architecture overview (Claude Help Center, verified 2026-10-12)" resource: https://support.claude.com/en/articles/14479288-claude-cowork-architecture-overview - id: anthropic-cowork-architecture-plans title: "Claude Cowork architecture overview, the plans it runs in the cloud on (Claude Help Center, updated 30 October 2026, verified 2026-10-31)" resource: https://support.claude.com/en/articles/14479288-claude-cowork-architecture-overview - id: anthropic-unified-enterprise title: "Set up and manage the new unified Claude experience for your Enterprise organization (Claude Help Center, verified 2026-11-03)" resource: https://support.claude.com/en/articles/17444462-set-up-and-manage-the-new-unified-claude-experience-for-your-enterprise-organization - id: anthropic-builtin-browser title: "Use the built-in browser in Claude Cowork (Claude Help Center, verified 2026-10-12)" resource: https://support.claude.com/en/articles/16607400-use-the-built-in-browser-in-claude-cowork - id: anthropic-docs-ga title: "Build live dashboards and animate explainers with Claude (Anthropic, 8 October 2026, verified 2026-10-12)" resource: https://claude.com/resources/articles/dashboards-and-motion - id: anthropic-artifacts title: "What are artifacts and how do I use them? (Claude Help Center, verified 2026-10-12)" resource: https://support.claude.com/en/articles/17153992-what-are-artifacts-and-how-do-i-use-them - id: anthropic-pricing title: "Claude pricing (Anthropic, verified 2026-10-12)" resource: https://claude.com/pricing - id: openai-chatgpt-work title: "ChatGPT is now a partner for your most ambitious work (OpenAI, 9 July 2026, verified 2026-10-12)" resource: https://openai.com/index/chatgpt-for-your-most-ambitious-work/ - id: openai-usage-limits title: "Understanding ChatGPT and Codex usage limits (OpenAI Help Center, verified 2026-11-04)" resource: https://help.openai.com/en/articles/20001625-understanding-chatgpt-and-codex-usage-limits - id: openai-dots title: "Introducing dots (OpenAI, 29 September 2026, verified 2026-10-22)" resource: https://openai.com/index/introducing-dots/ - id: openai-pro-tiers title: "About ChatGPT Pro tiers (OpenAI Help Center, verified 2026-10-12)" resource: https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers - id: anthropic-claude-in-chrome title: "Get started with Claude in Chrome (Claude Help Center, verified 2026-10-17)" resource: https://support.claude.com/en/articles/12012173-get-started-with-claude-in-chrome - id: anthropic-m365 title: "Work across M365 apps (Claude documentation, verified 2026-10-17)" resource: https://claude.com/docs/office-agents/work-across-apps - id: anthropic-google-workspace title: "Claude now works with Google Docs, Sheets, and Slides (Anthropic, 6 October 2026, verified 2026-10-13)" resource: https://claude.com/resources/articles/claude-now-works-in-google-docs-sheets-and-slides - id: openai-excel-sheets title: "ChatGPT for Excel and Google Sheets (OpenAI Help Center, verified 2026-10-13)" resource: https://help.openai.com/en/articles/20001063 - id: openai-work-files title: "Creating and editing documents, spreadsheets, and presentations with ChatGPT Work (OpenAI Help Center, verified 2026-10-13)" resource: https://help.openai.com/en/articles/20001278-creating-and-editing-documents-spreadsheets-and-presentations-with-chatgpt-work - id: anthropic-scheduled-tasks title: "Schedule recurring tasks in Claude Cowork (Claude Help Center, verified 2026-11-03)" resource: https://support.claude.com/en/articles/13854387-schedule-recurring-tasks-in-claude-cowork - id: anthropic-file-uploads title: "Upload files to Claude (Claude Help Center, verified 2026-10-14)" resource: https://support.claude.com/en/articles/8241126-upload-files-to-claude - id: openai-desktop-app title: "ChatGPT desktop app (ChatGPT documentation, verified 2026-10-20)" resource: https://learn.chatgpt.com/docs/app - id: openai-scheduled-tasks title: "Scheduled tasks walkthrough (ChatGPT documentation, verified 2026-10-20, with the Run on this computer setting and the web's text box seen in the apps on 2026-10-20)" resource: https://learn.chatgpt.com/training/walkthroughs/scheduled-tasks - id: anthropic-duet title: "Claude Duet (Anthropic, 21 October 2026, verified 2026-10-31)" resource: https://claude.com/resources/articles/claude-duet - id: opencode-v2 title: "OpenCode v2 documentation (OpenCode, 2.0 announced 12 October 2026, verified 2026-10-22)" resource: https://opencode.ai/v2/docs/ - id: opencode-daemon title: "OpenCode 2.0 Turned Your Coding Agent Into a Daemon (Towards AI, 16 October 2026, verified 2026-10-22)" resource: https://pub.towardsai.net/opencode-2-0-turned-your-coding-agent-into-a-daemon-4663135862c4 - id: opencode-cli title: "CLI, the background service (OpenCode v2 documentation, verified 2026-10-23)" resource: https://opencode.ai/v2/docs/cli/ - id: openclaw-proactive-check title: "Heartbeat, now the proactive check job (OpenClaw documentation, verified 2026-10-23)" resource: https://docs.openclaw.ai/gateway/heartbeat - id: openclaw title: "OpenClaw README (OpenClaw Foundation, GitHub, MIT, verified 2026-10-22)" resource: https://github.com/openclaw/openclaw - id: openclaw-security title: "OpenClaw: From Viral Prototype to Agentic Infrastructure (Catalaize, February 2026, verified 2026-10-22)" resource: https://catalaize.substack.com/p/openclaw-from-viral-prototype-to - id: omnara title: "Omnara README (Omnara, GitHub, Apache 2.0, verified 2026-10-22)" resource: https://github.com/omnara-ai/omnara/blob/main/README.md generated: at: 2026-11-05T06:58:08Z by: process:panaversity trust_tier: unverified build_id: sha256:46fd53c5d7ab1c14f6347908d45c9fdaecda4d914439923ee878bdcecd57803e dirty: true ksor_version: 0.0.60 --- **In everyday life.** Two banks both offer mobile check deposit. One puts it inside the main app. The other has a separate deposit app. The idea is the same, the packaging differs, and you check each bank's rules before relying on them. Some people also use a free, open app that works with any bank. It gives them more control, and more to check themselves. The principle so far names no product. Here is how the two AI vendors this book covers, and the main open-source tools, put concepts 1.2 to 1.4 into practice today. > **Anthropic.** > > **One conversation.** Claude chat and Claude Cowork are now one > conversation, with no mode to pick.[^anthropic-one-claude-post] Claude > decides whether a request needs a quick answer or a task. Anthropic is > moving accounts over in stages, so some get this before others, even on > the same plan. If your message box still shows Chat and Cowork options, > your account has not moved over yet.[^anthropic-one-claude-nov] > **Where you open it.** The same Claude runs in three places: in your > web browser at claude.ai, in the Claude desktop app for Mac, Windows > and Linux, and in the Claude mobile app. The Linux app is in beta, for > Ubuntu 22.04 or later and Debian 12 or later. On other Linux systems, > Anthropic points to Claude Code, its coding agent, in the terminal > instead.[^anthropic-desktop-linux] Your conversations, tasks and > artifacts are saved to your account, so you can start on one and > pick up on another. For everyday work, > use whichever you prefer. Install the desktop app when a task needs > your own computer: it is the bridge to your local folders, described > under *Your computer* below. > > ![One text box with no mode to pick. Claude reads what you typed: a > question gets an answer seconds later in the same chat; a request that > describes an outcome becomes a task that keeps running with the laptop > closed and ends in finished work saved as an > artifact.](./img/one-conversation.png) > > **Where tasks run.** On the Pro, Max and Ultra plans, tasks run in the > cloud: they keep going after you close the laptop, and scheduled tasks > run while your computer is off.[^anthropic-one-claude-plans] Since October > 6, 2026, new tasks on these plans run there. A task you started on your > computer before that date stays there until it is done.[^anthropic-cowork-surfaces] > > ![Two timelines. A task you hand off: the brief goes in at 5:00 pm, the > laptop closes at 5:10, and the task finishes in the cloud at 5:40 — the > closed laptop makes no difference. A scheduled task: the computer is shut > down at 11:00 pm, the schedule fires at 7:00 am while it is off, and the > finished result is waiting when it is turned on at > 9:00.](./img/where-tasks-run.png) > > **The sandbox.** Each cloud session runs in its own isolated, temporary > sandbox on Anthropic's servers, separate from every other session's. It > cannot reach your home or company network. It never holds your passwords: > the only keys inside are temporary passes, called session tokens, that > expire within hours. When the session ends, the sandbox is destroyed. > This is true for every kind of task. A task you hand off has one sandbox > for its session. A scheduled task runs as its own session each time it > fires, so each run gets a fresh sandbox — and between runs there is no > machine at all, only the task's saved instructions and its run > history.[^anthropic-scheduled-tasks] Anthropic's setup page for Enterprise > owners adds a detail, as of 3 November 2026. After a few idle minutes, the > sandbox is stopped and its disk is saved. It expires 30 days after it was > created.[^anthropic-unified-enterprise] No sandbox is permanent. What is > permanent lives in your account: the brief, the sessions, the artifacts. > Calls to your connected apps are made from Anthropic's servers, outside > the sandbox.[^anthropic-cowork-architecture] > > ![The life of a sandbox in three stages. It starts as a fresh, empty > workspace holding no passwords — only temporary passes, called session > tokens, that expire within hours; a scheduled task gets a fresh sandbox > each run. While it works — on copies of your files, never the originals — > it cannot reach your home or company network or other sessions' > sandboxes, and app calls are made for it by Anthropic's servers from > outside. When the session ends it is destroyed — wiped, not archived — > with your results and your uploads already safe outside > it.](./img/sandbox-lifecycle.png) > > **Your computer.** Claude Desktop is the bridge to your own machine. A > cloud task can reach only the folders you connected, and only while > Desktop is open on that computer. If Desktop is closed, the task keeps > running but cannot reach those folders.[^anthropic-cowork-surfaces] As of > 3 November 2026, two Anthropic pages differ on a scheduled task that needs > files or apps on your computer. The scheduling page says it runs only on > your computer.[^anthropic-scheduled-tasks] The page on web, desktop and > mobile says new scheduled tasks run in the cloud.[^anthropic-cowork-surfaces] > So check where your task says it runs. When a task > needs a website, Desktop gives it a built-in browser in a side panel, > separate from your own browser.[^anthropic-builtin-browser] Claude in > Chrome, a browser extension on paid plans, works in your own Chrome > instead: it reads, clicks and navigates websites alongside you. Its side > panel is in beta.[^anthropic-claude-in-chrome] > > ![You send a brief, plus copies of any files you attach, and close the > laptop. The task runs in a temporary sandbox in Anthropic's cloud that > cannot reach your network; results are saved to your durable account; > connected apps are called from Anthropic's servers outside the sandbox; > and your computer is reachable only through the Claude Desktop bridge, > for connected folders, while Desktop is > open.](./img/cloud-task-sandbox.png) > > **Inside your files.** Claude's own editors for documents, slide decks and > designs are built into the Claude app itself — no other software needed; > the next note, on where results are kept, covers them. For the files you > already keep in Google, Claude for Google Workspace, in public beta on > paid plans, adds Claude as a sidebar inside Google Docs, Sheets and > Slides. It reads the open file, sees what you have selected, and changes > the file in place. In Docs, bigger rewrites arrive as cards you apply or > dismiss one at a time; in Sheets it writes formulas, pivot tables and > charts; in Slides it builds new slides from the deck's own layouts, then > flags overlapping or unreadable elements. Separate Docs, Sheets and > Slides connectors run the other direction: you stay in Claude, paste a > file link, and Claude creates or edits the Google file from the > chat.[^anthropic-google-workspace] For files in Microsoft 365, Claude for > Excel, PowerPoint, Word and Outlook are add-ins on paid plans. With > cross-app mode on, Claude reads from one open file and changes another, > such as numbers from an Excel model put into a PowerPoint slide. It works > only with files that are already open.[^anthropic-m365] > > ![Two homes for your files. Files made in Claude live in its built-in > editors — Docs, Slides and Design, with Dashboards and Motion in beta — > and are kept in the Artifacts section. Files kept in Google get Claude > brought to them: a sidebar inside the open file proposing edits you apply > or dismiss, or, in the other direction, Claude editing the Google file > from the chat when you paste its link.](./img/inside-your-files.png) > > **Where results are kept.** Your sessions and files are saved to your > account, so the same work appears on web, desktop and mobile. The Claude > Docs, Claude Slides and Claude Design editors are out of beta and on every > plan, including Free; two newer editors are in beta — Claude Dashboards, > which builds live dashboards, on paid plans, and Claude Motion, which > builds short animated explainers, on Team and Enterprise > plans.[^anthropic-docs-ga] There is no spreadsheet editor in this lineup > yet. Spreadsheet work comes back as a file you download, or happens in > Google Sheets through the sidebar — and for live numbers from connected > data, the Dashboards beta covers part of what a spreadsheet > would.[^anthropic-docs-ga] The catalog of editors follows a pattern — > each kind of work people keep and share eventually gets its own — so > expect this list to grow, and read each addition with this chapter's > four questions. Your team and Claude can edit the same doc, deck or > design together, in the same file.[^anthropic-docs-ga] Everything you > make is called an artifact. All your artifacts, from all your chats, are > collected in one place: the Artifacts section in the Claude sidebar. > Close a chat today, and what Claude built for you in it is still there > next month.[^anthropic-artifacts] > > ![Three chats — Monday's, Wednesday's, and one from last month — each > produce an artifact. All of them are collected in the Artifacts section > of the Claude sidebar, the same on web, desktop and mobile, still there > after each chat is closed. Plain files are different: an image, a data > file or a PDF arrives as a file card and stays in the chat that > delivered it — reopen the chat to download it > again.](./img/artifacts-one-place.png) > > Not everything Claude makes is an artifact. A task also produces plain > files — an image, a data file, a PDF, a zip. These are handed to you in > the conversation itself: each arrives as a file card in the chat, and > because the chat is saved to your account, you can reopen it and download > the file again from any device.[^anthropic-cowork-surfaces] A working > file the task made but never handed over is different: it lived only in > the sandbox, and it is wiped with it — so if a task mentions a file you > want, ask for it before you move on. And when you want a result on your > own computer, a task can write it into a connected folder, through the > bridge described above.[^anthropic-cowork-surfaces] > > The files you attach follow the same rule in reverse. An upload belongs > to the chat you attached it to: reopen the chat, and it is still there. > A task works on a copy of your upload inside the sandbox, so your > original is untouched when the sandbox is wiped. And for files you will > use across many chats — a price list, a policy, a product catalog — > there is a better home than any single chat: a project's own Files > section, which [Chapter 2](../what-is-an-ai-worker/worker-runtime-channel.md) introduces.[^anthropic-file-uploads] > > Work done through the Google Workspace sidebar is the exception: those > results live in the Google file itself, in your > Drive.[^anthropic-google-workspace] > > **A worker that keeps going.** Claude Duet is Anthropic's standing > worker. It is a research preview, with a waitlist for US Pro, Max and > Ultra subscribers. Duet is separate from the main Claude app for now. You reach > it in one ongoing thread, typed or spoken, in its own app on iOS and > Android or on claude.ai. It starts with the memory, skills and connected > apps from your Claude account. It has its own virtual machine with its own > browser, so your laptop and phone can be off. Each morning it sends a short > brief. During the day it checks your connected accounts every few hours, > and it messages you only when something needs attention. It reads, drafts > and prepares work on its own. Before it sends a message, makes a payment, > signs a document or deletes anything, it shows exactly what it will do and > waits for your approval. Some actions, such as trading securities, are not > allowed even with approval. Its background checks count toward your plan's > usage limits. The page does not say how long its machine, or the files on > it, last.[^anthropic-duet] > > **Status and price.** A permission setting in the message box decides how > often Claude asks. As of 3 November 2026, Auto, the default on Pro, Max, > Ultra and Team, keeps working, with automated safety checks. Manual, the > default on Enterprise, asks before each action.[^anthropic-one-claude-nov] > Anthropic's launch post says Claude asks by default, so read what your own > message box shows.[^anthropic-one-claude-post] Cloud tasks are available on the Pro, > Max and Ultra plans.[^anthropic-cowork-architecture-plans] On Team plans > they are in beta on web and mobile, and on Enterprise plans an owner turns > them on.[^anthropic-cowork-architecture] > Handing off and scheduling tasks starts with the Pro plan, at $20 a month > billed monthly.[^anthropic-pricing] > **OpenAI.** > > **One conversation, and dots.** ChatGPT is coming together in one > conversation too. OpenAI says it is bringing what Work, its agent for longer > tasks, can do into the main ChatGPT experience.[^openai-usage-limits] Beside > that conversation, OpenAI runs dots. > > **If your app still shows a switch.** As of 4 November 2026, OpenAI's page on > the desktop app still describes a switch between Chat and > Work.[^openai-moving-desktop] If you see it, choose Work for a task. From > December, OpenAI says Plus, Pro and Business plans will share one allowance > across ChatGPT and Codex.[^openai-usage-limits] > > **Tasks in ChatGPT.** In a task, an agent acts across your apps and files. It can > keep working on a project for hours, and it turns a goal into finished > slides, spreadsheets, documents and websites. It offers scheduled tasks, a > built-in browser, and local files and apps through the desktop > app.[^openai-chatgpt-work] > > **Scheduled tasks.** Scheduled tasks come with this agent, but you > make them from their own item in the left menu, Scheduled, and a task > can also be scheduled from inside a chat.[^openai-desktop-app] Where it > runs depends on how you make it. On the web, you describe the task and > its timing in the text box, and it works with your connected tools. In > the desktop app, a form sets the timing, and its advanced settings > include Run on this computer. Turned on, the task runs on your computer, > which must be on with the app running. Turned off, it runs in the cloud, > where OpenAI offers it. A cloud run cannot reach the folders on your > computer.[^openai-scheduled-tasks] > > **Inside your files.** The ChatGPT add-on for Microsoft Excel and Google > Sheets puts a sidebar inside the open spreadsheet that reads and edits > the live cells.[^openai-excel-sheets] Through its Google Workspace apps, > ChatGPT can also create and edit native Google Docs, Sheets and Slides > from the chat, so the file it produces is a real file in your Drive, not > an attachment in a conversation.[^openai-work-files] > > **Dots.** A dot is an always-on agent. Each dot has its own cloud computer > and browser, which you can open to check its work. With your permission, > it can also use your laptop. A dot keeps working between conversations, > connects to apps through plugins, and carries context across ChatGPT, > Slack and Teams.[^openai-dots] > > ![A dot at the center of its world. It works on its own cloud computer > and browser, which you can open any time to check; it reaches you in > ChatGPT, Slack and Teams with the same identity and context; it connects > to your apps through plugins; and it can use your laptop only with your > permission. Between your conversations it keeps working — you chat on > Monday, and on Thursday it remembers.](./img/dots-anatomy.png) > > **Rules and oversight.** During what OpenAI calls proactive research, a > dot uses only read-only tools. Custom Rules set what it may do alone, what > needs approval and what is blocked. Activity View shows its > work.[^openai-dots] > > ![An illustrated example of a dot's controls. Custom Rules sort its > actions into three tiers: things it may do alone, such as reading an > inbox; things that need your approval, such as sending an email; and > things that are never allowed. Activity View logs every action it took, > including work waiting for approval. During proactive research — work the > dot starts on its own — it uses read-only tools: it can look, not > touch.](./img/dot-rules.png) > > **Status and price.** Dots are rolling out to Pro and Business Premium > users in eligible markets. Enterprise users can try a beta when their > workspace admin turns it on. Each user starts with one primary dot, > included in the plan. Chatting with your dot does not count toward your > usage limits, but tasks it starts for you do.[^openai-dots] > ChatGPT Pro comes in three monthly tiers, at $100, $200 and > $500.[^openai-pro-tiers] > **Open source.** > > **What changes.** Open-source tools are free programs whose code anyone > can read. You install them on your own computer or server, and you choose > the AI model, including one you run yourself. No vendor runs them for > you, so there is no vendor sandbox and no vendor help center. You take on > the setup, the updates and the security. There is also no single app that > decides between answering and working. You pick the tool for the job. > This book uses three projects. Two belong in this chapter. The third, > Omnara, a runtime for teams and products that you can host yourself, > comes in later chapters.[^omnara] > > **OpenCode, for a task.** OpenCode is an open-source coding agent, the > counterpart of Claude Code and Codex, the coding agents from Anthropic and > OpenAI. [Chapter 12](../../building-ai-workers/agentic-coding/overview.md) covers all three. Version 2.0 runs as one shared > background service on your machine. Its terminal app, desktop app, web interface and > your own scripts all connect to that one service.[^opencode-cli] There is no hosted web > app to sign in to: even the web interface connects to the agent on your > own machine. It works with many model providers, including local > models.[^opencode-v2] Like a Claude task, > an OpenCode task is briefed, runs and ends. But it runs on your computer, > not in a vendor's cloud, so the files it makes are in your own folders. > One report points out a new risk: because the agent now runs outside the > terminal, a sandbox around the terminal no longer contains > it.[^opencode-daemon] > > **OpenClaw, for ongoing work.** OpenClaw is a standing assistant that you > host yourself. A scheduled job, the proactive check, wakes the agent > every 30 minutes by default and sends what it finds to you in a direct > message.[^openclaw-proactive-check] You reach it through apps you already use, such as WhatsApp, > Telegram, Slack and Discord. It is free under the MIT license and run by > the OpenClaw Foundation, an independent nonprofit.[^openclaw] > > **Where things run, and where they are kept.** The worker, its memory and > its passwords stay on hardware you control. Your prompts still go to the > company that runs the model, and to the chat apps you use, unless you run > the model yourself. OpenClaw's tools run directly on your computer unless > you set up a sandbox.[^openclaw] So there is no disposable sandbox unless > you build one. > > **Security.** Take this part seriously. An early security report found over > 30,000 OpenClaw copies open to the internet, many without passwords, and > said security firms called its default settings unsafe.[^openclaw-security] > Its current guidance is better. It treats incoming messages as untrusted, > and unknown senders must pair with it before they can message it. Read > each project's security guide. Run a standing assistant on a separate > machine, with its own accounts and the smallest possible > permissions.[^openclaw] > > **Status and price.** The software is free. You pay for the model you > use, unless you run one on your own hardware. OpenCode also offers > OpenCode Go, an optional $10-a-month plan for coding > models.[^opencode-v2] Open-source tools are not on Anthropic's > certification exams. **The text box moves into the file.** Until now we have treated the text box as a separate place: you go to the AI, give it work, and carry the results back. Both vendors have closed that gap — the box now also opens inside the file you are already working in. Notice what this does to the three ideas this chapter keeps separate. When the AI edits the file you have open, *where the result is kept* stops being a trap: the result is the file, in your Drive, where you already look for it. Maria's lost spreadsheet could not have happened this way. And the suggestion cards — change proposed, human applies or dismisses — are a small preview of the working rhythm this book builds on in [Chapter 3](../the-10-80-10-operating-rhythm/one-rhythm.md): the AI does the work, and you stay the one who approves it. ![Before: your file and the AI chat are two windows, and you carry work in and copy results back by hand. Now: the AI opens as a sidebar inside the file, reads it, and proposes edits as cards you apply or dismiss — the result is the file itself.](./img/text-box-into-file.png) **The differences that change a decision.** Six matter. 1. **Where the agentic work starts.** In Claude and in ChatGPT, it starts in one conversation. Anthropic says Claude decides whether a request needs a quick answer or a task.[^anthropic-one-claude-post] OpenAI has not said how ChatGPT makes that choice, so read what your brief gets back. A standing worker has its own place. Duet has its own app,[^anthropic-duet] and you pick a dot by name. A dot can also start tasks for you.[^openai-dots] With open-source tools, you install the tool first, then pick it by name: OpenCode for a task, OpenClaw for ongoing work. 2. **How continuity is provided.** A Claude task is briefed, runs, and ends, and so does an OpenCode task. When the task is done inside a defined role with an owner, this is how a **Delegated AI Worker** does its work: a worker briefed for one defined job, with a named owner. Claude also reaches recurring work through schedules and routines ([Chapter 9](../../managing-ai-workers/work-that-runs-without-you/scheduled-and-event-triggered.md)). Duet, a dot and OpenClaw each keep their identity and context between conversations and take initiative. That brings them close to a **Standing AI Worker**, one with a lasting identity, ongoing responsibilities and its own initiative. [Chapter 2](../what-is-an-ai-worker/anatomy-of-an-ai-worker.md) defines both kinds. In every case, continuity comes from the worker's identity and context. A machine that stays on only runs the worker. The context is what lets it pick up where it left off. That context, such as a standing worker's memory or a conversation's history, is itself durable state: an AI's record of you and the work so far. What it should hold, where it lives, and when to trust it is [Chapter 8](../../managing-ai-workers/context-memory-knowledge-and-state/memory-and-ssor.md)'s subject. For now, ask of it what you ask of anything kept: where is it, and who can reach it? 3. **What you must verify yourself.** Anthropic documents its session sandbox.[^anthropic-cowork-architecture] But neither Anthropic's Duet page nor OpenAI's launch post says how long the worker's own computer, or the files on it, last.[^anthropic-duet][^openai-dots] With open-source tools, you can read the code, but nobody checks your setup for you. Check each AI vendor's help center, and each project's security guide, before you rely on such details. 4. **What comes back.** Claude's agentic work is built around an outcome. A task ends in finished files, documents, slide decks or designs, made in editors inside the same Claude app on web, desktop and mobile.[^anthropic-one-claude-post] A dot is a lasting co-worker in the background. It works across your apps, reaches you in ChatGPT, Slack and Teams, and starts tasks when the job needs finished materials.[^openai-dots] Duet works the same way in the background, in one ongoing thread, and waits for your approval before it acts outward.[^anthropic-duet] OpenCode's results are files in folders on your own machine. OpenClaw reaches you in the chat apps you already use.[^openclaw] 5. **What it costs to start.** Claude's tasks start on the Pro plan, at $20 a month.[^anthropic-pricing] Duet is for Pro, Max and Ultra subscribers, and for now only through a waitlist in the US.[^anthropic-duet] A dot needs ChatGPT Pro, from $100 a month, or a Business Premium plan, and Enterprise plans have a beta.[^openai-pro-tiers] Open-source tools are free to install. You still pay for the model, unless you run one on your own hardware, and you pay in your own time for setup and security.[^opencode-v2] Prices change, so check each pricing page on the day you decide. 6. **Who holds your data.** With Anthropic and OpenAI, your sessions, your files and the worker's memory live in the vendor's cloud, under the vendor's terms. With open-source tools, the worker's state, memory and passwords stay on hardware you control. Your prompts still go to the company that runs the model, unless you run the model yourself.[^openclaw] Control moves to you, and so does the security work. One contrast underlies the first three differences, and it is worth saying plainly. **For a delegated task in Claude, the machine is always disposable and the record is always durable.** Sandboxes are born for a run and destroyed, while the brief, the history and the artifacts live on in your account. ChatGPT tasks are delegated too. **A standing worker stands that on its head: the machine itself is the standing thing.** A dot has its own cloud computer, and Duet its own virtual machine. The files on them live wherever they do, for a lifespan neither launch page states. Open-source tools move the machine onto your own hardware, for a delegated worker such as an OpenCode task and for a standing one such as OpenClaw. It lasts as long as you keep it, and so does everything on it, including the passwords you gave it. So the question to carry out of this comparison is: *what survives, the machine or the record?* Keep what you care about where the survivor is. If a task without a permanent machine looks like a gap to you, hold the thought. [Chapter 4](../the-architecture-in-one-picture/worker-is-the-records.md) argues it is a design choice, and shows what takes its place. ![What persists, the workspace or the record? A grid with two kinds of worker, delegated and standing, and two places the work runs, a vendor's cloud and your own hardware. Delegated, in a vendor's cloud: a Claude task or routine gets a fresh workspace for each run, removed afterward, while the brief, run history, sessions and artifacts stay in your account. What persists is the record, not the workspace. ChatGPT tasks are delegated too. Delegated, on your own hardware: an OpenCode task, run by a coding agent, runs on your computer, with no hosted web app. Its terminal, desktop app and web interface all connect to the agent on your machine. What persists is your folders, as long as you keep them, so back them up. Standing, in a vendor's cloud: Claude Duet and an OpenAI dot each have a cloud computer of their own that keeps running between conversations. What persists is the machine and its files, between conversations. Neither launch page states a retention period. Standing, on your own hardware: OpenClaw runs on a machine you own and keep on. Everything on it persists, passwords too, so use a separate machine, its own accounts and the smallest permissions. Keep what you care about where the survivor is.](./img/machine-or-record.png) **Which surface for which job.** The choice follows the ladder. On both AI vendors, the first two rungs sit in one conversation, so your brief says which you want. Ask a question, or ask for a draft or an explanation you will read once. Give a clear outcome when a finished result should come back, such as the invoice register. A schedule is not a third choice. It repeats a task: the work runs the same way, but a clock starts it instead of you. For ongoing work that should continue between conversations, such as keeping track of a vendor inbox, use a standing worker. On OpenAI's side that is a **dot**. Set its Custom Rules first, so you decide what it may do alone. On Anthropic's side it is Duet, once it reaches your account. In open source, OpenCode answers questions and does tasks with a clear outcome. OpenClaw handles ongoing work, but only on a separate machine you have secured. [Chapter 2](../what-is-an-ai-worker/both-ai-vendors.md) adds OpenAI's Space, where a team keeps the documents it edits together, and completes the comparison. ![Three kinds of job: a question, a clear outcome, and ongoing work. On Anthropic's side, a question or an outcome is handled in one conversation, where your brief decides and a schedule repeats a task. Claude Duet, in preview, handles ongoing work and asks before it acts outward. On OpenAI's side, a question or an outcome is also handled in one conversation, where you give a brief and a schedule repeats a task. A dot, with its Custom Rules set first, handles ongoing work. In open source you install the tool, then pick it: OpenCode for questions and tasks on your own machine, and OpenClaw for ongoing work, only on a separate, secured machine. Define the job first, then choose how to delegate it.](./img/which-surface.png) **The invariant.** Products will keep changing, but on any runtime four things stay true. First, every request is routed one of two ways: to answering, or to working. Sometimes the product routes it, and sometimes you do, by picking the tool. Second, the work runs in an execution environment, and you must know who controls it. It may be a vendor's sandbox or cloud computer, or your own machine that you must secure yourself. Either way, your prompts go wherever the model runs. Third, anything you want to keep must be saved to a known place, and you must understand who can reach it and how long it stays. And fourth, that durable state, such as your files, your artifacts or a standing worker's memory, is still not the company's governed record, the system your organization treats as the official truth. That last distinction is the seed of this book's architecture, and [Chapter 4](../the-architecture-in-one-picture/overview.md) makes it precise. ![Four distinctions as a pipeline. First, route the request to answering or to doing work. The product routes it, or you do, by picking the tool. Second, know who controls where the work runs: a vendor's sandbox, cloud computer or remote browser, or your own machine, which you must secure yourself. Either way, prompts go wherever the model runs. Third, name where what you keep is saved, whether files, artifacts or a worker's memory, and ask who can reach it and how long it stays. Fourth, saving does not automatically make it the governed record, the system your organization treats as the official record. Saved output is not governed truth. Chapter 4 draws the architecture that closes that gap.](./img/the-invariant.png) [^anthropic-one-claude-post]: Claude Cowork and chat are now one Claude, Anthropic, 16 September 2026. [^anthropic-one-claude-plans]: Claude Cowork and chat are one Claude, Claude Help Center, updated 30 October 2026. [^anthropic-one-claude-nov]: Claude Cowork and chat are one Claude, Claude Help Center, accessed 3 November 2026. [^anthropic-cowork-surfaces]: Use Claude Cowork on web, desktop, and mobile, Claude Help Center. [^anthropic-desktop-linux]: Claude Desktop on Linux (beta), Claude Code documentation. [^anthropic-cowork-architecture]: Claude Cowork architecture overview, Claude Help Center. [^anthropic-unified-enterprise]: Set up and manage the new unified Claude experience for your Enterprise organization, Claude Help Center. [^anthropic-cowork-architecture-plans]: Claude Cowork architecture overview, Claude Help Center, updated 30 October 2026. [^anthropic-scheduled-tasks]: Schedule recurring tasks in Claude Cowork, Claude Help Center. [^anthropic-builtin-browser]: Use the built-in browser in Claude Cowork, Claude Help Center. [^anthropic-docs-ga]: Build live dashboards and animate explainers with Claude, Anthropic, 8 October 2026. [^anthropic-google-workspace]: Claude now works with Google Docs, Sheets, and Slides, Anthropic, 6 October 2026. [^anthropic-claude-in-chrome]: Get started with Claude in Chrome, Claude Help Center. [^anthropic-m365]: Work across M365 apps, Claude documentation. [^anthropic-artifacts]: What are artifacts and how do I use them?, Claude Help Center. [^anthropic-file-uploads]: Upload files to Claude, Claude Help Center. [^anthropic-pricing]: Claude pricing, Anthropic. [^openai-chatgpt-work]: ChatGPT is now a partner for your most ambitious work, OpenAI, 9 July 2026. [^openai-usage-limits]: Understanding ChatGPT and Codex usage limits, OpenAI Help Center. [^openai-excel-sheets]: ChatGPT for Excel and Google Sheets, OpenAI Help Center. [^openai-work-files]: Creating and editing documents, spreadsheets, and presentations with ChatGPT Work, OpenAI Help Center. [^openai-dots]: Introducing dots, OpenAI, 29 September 2026. [^openai-pro-tiers]: About ChatGPT Pro tiers, OpenAI Help Center. [^openai-desktop-app]: ChatGPT desktop app, ChatGPT documentation. [^openai-scheduled-tasks]: Scheduled tasks walkthrough, ChatGPT documentation. [^anthropic-duet]: Claude Duet, Anthropic, 21 October 2026. [^opencode-v2]: OpenCode v2 documentation, OpenCode. [^opencode-daemon]: OpenCode 2.0 Turned Your Coding Agent Into a Daemon, Towards AI, 16 October 2026. [^openclaw]: OpenClaw, OpenClaw Foundation, README on GitHub. [^openclaw-security]: OpenClaw: From Viral Prototype to Agentic Infrastructure, Catalaize, February 2026. [^opencode-cli]: CLI, OpenCode v2 documentation. [^openclaw-proactive-check]: Heartbeat, OpenClaw documentation. [^omnara]: Omnara, README on GitHub. [^openai-moving-desktop]: Moving to the new ChatGPT desktop app, OpenAI Help Center.