Lesson 2.52.5 How today's tools realize it

In everyday life. Two phone companies both sell small, medium and large plans, but the names and limits differ. You compare what each plan does for you, not what it is called.

Anthropic

Anthropic, as verified 12 October 2026.

Chat and tasks. One Claude conversation routes between answering and tasks (Chapter 1).¹² As of 31 October 2026, Team and Enterprise plans keep chat and Cowork separate unless the organization has turned on the one conversation, which is in beta on Team.³ A task is briefed, runs and ends. Claude Tag, in beta on Team and Enterprise, is one shared Claude in a team's Slack channel. It takes on work when someone tags it in a team conversation, and remembers context across days.⁴

Recurring and continuing work. For recurring work, Claude offers scheduled tasks on paid plans. They run remotely, even while your computer is asleep.⁵ As of 3 November 2026, two Anthropic pages differ on a task that needs files or apps on your computer. The scheduling page says it runs only on your computer.⁵ The page on web, desktop and mobile says new scheduled tasks run in the cloud, and a cloud session reaches your files only while the Claude Desktop app is open. If the app is closed, the session keeps running but cannot reach them.⁶ So check where your task says it runs. Each run starts fresh and picks up from the records (Chapter 1) — the second way to continuity in Figure 2.5. Chapter 9 compares scheduled tasks with an always-on agent.

A standing worker, verified 31 October 2026. Claude Duet is Anthropic's standing worker. It keeps working after you close its app. That is the always-on way to continuity in Figure 2.5. It is a research preview with a waitlist, for Pro, Max and Ultra subscribers in the US. For now it has its own app, separate from the main Claude app. It has a virtual machine with its own browser, so your laptop and phone can be off. It checks your connected accounts every few hours. How long its machine and the files on it last is not stated.⁷

Projects and memory. Projects hold standing instructions and a knowledge base of files for one area of work, on every plan. Team and Enterprise members can share them.⁸ Memory saves facts as topics you can read, edit or delete in Settings, and each project has its own. As of 31 October 2026, it is on by default on Free, Pro, Max and Ultra; on Team and Enterprise, owners decide.⁹

Connectors and skills. Connectors link Claude to other apps. The Google Workspace connectors reach Gmail, Google Calendar and Google Drive. By default, Claude asks you before it sends an email or shares a file.¹⁰ Skills are folders of instructions, scripts and resources that Claude loads when a task needs them. Anthropic provides some, such as skills for Word and Excel files, and you can add your own. Skills follow an open standard, so they are not only for Claude.¹¹

Code execution and research. Code execution and file creation, a setting that is on by default, lets Claude write and run code in a sandbox. It works out numbers and builds spreadsheets, documents and other files.¹² Skills need this setting turned on.¹¹ Research, on paid plans, runs several searches that build on each other, across the web and connected apps, and returns citations.¹³

Artifacts and Claude Docs. Artifacts open beside the conversation as documents, slide decks, designs, dashboards or small tools that you can edit, return to and share. You can ask for one in any chat, or start one from a template. The Docs, Slides and Design templates are on every plan, including Free.¹⁴¹⁵ Dashboards are in beta on paid plans, and Motion, for animations, is in beta on Team and Enterprise.¹⁵ In Claude Docs, a team writes a document together with Claude in real time.¹⁶

Models. Claude's models each have a job.¹⁷

• Haiku, the fastest and lowest-cost, for high-volume steps, such as sorting documents or pulling data out of them.¹⁸
• Sonnet, for speed and capability in everyday work, such as writing and data analysis.
• Opus, for complex agent work and enterprise work. Anthropic's guide for developers says most workloads start with it.
• Fable, the most capable model open to all customers, for work that runs for hours, such as research that ends in a finished report.

The current versions are Haiku 5.5, Sonnet 5.5, Opus 5.5 and Fable 5.1. Anthropic's models overview lists every model, with its price for developers.¹⁸ One model menu next to the send button sets the model and its effort level, and you can change both at any point.¹⁹ Anthropic advises that tuning effort often works better than switching models.¹⁷

For the AP Worker, that could mean Haiku to sort each night's incoming invoices, Sonnet to draft replies to vendor questions, and Opus for the weekly payment-run proposal. Fable comes in only if a long job, such as checking every vendor contract at year end, still falls short on Opus at a higher effort.¹⁸ Test each choice on the worker's own cases first (2.4).

OpenAI

OpenAI, as verified 6 to 15 October 2026.

The conversation and dots are described in Chapter 1. Quick questions and longer tasks come together in one conversation. Work, OpenAI's name for its agent, does longer, multi-step work that delivers documents, spreadsheets, presentations, reports and Sites.²⁰ A dot keeps working between conversations — the always-on way to continuity in Figure 2.5. A team can also reach ChatGPT in its own channels, as @ChatGPT in Slack and Microsoft Teams.²¹

Specialist dots, verified 23 October 2026. Your own dot works on your behalf. A specialist dot takes on dedicated responsibilities inside a company. The company sets up each one with its own identity, its own logins and access to the systems it needs. OpenAI is starting with focused enterprise pilots, which are early trials with companies. In each one, its engineers work with the company to define the dot's responsibilities, its tools, and how people review and approve its work.²² With its own identity and dedicated responsibilities, it is the nearest thing in this section to a Standing AI Worker built with a vendor's help.

Projects and memory. Projects keep related chats, files and instructions together, and can be shared.²³ From a project you can start a task that uses the project's context.²⁰ Memory keeps saved memories that you can review, correct and delete, and a second setting uses past chats.²⁴

Apps and skills. In ChatGPT, connectors are called apps.²⁵ A skill packages instructions, resources and optional scripts, so ChatGPT or Codex can follow a workflow reliably. Skills follow the same open standard as Claude's.²⁶

Deep research. Chosen from the tools menu, it proposes a research plan you can edit, then returns a report with citations.²⁷

Space. Space is a home for pages and files.²⁸ Collaborators edit the same page, each with their own ChatGPT. Sharing a page does not give other people access to your private chats or memory, but anything written on the page is visible to everyone who can view it. Space is on Pro, Business and Enterprise plans.²⁹

Models. Where your app still shows Chat and Work, each has its own model picker. On Business plans, you choose a reasoning level in chat, from Instant to Extra High, or Pro.³⁰ Work and Codex, OpenAI's coding agent, use their own models, with a Default option that sets the model and reasoning level together.²⁰ OpenAI's page on ChatGPT Business models and limits lists the models and their usage limits.³⁰

Open source

Open source, as verified 23 October 2026.

OpenCode, for a task. OpenCode is an open-source coding agent. It runs on your own computer, and so do its tools.³¹ It is an agent that works on its own: you brief one task, it runs, and it ends. That suits a Delegated AI Worker.

OpenClaw, for ongoing work. OpenClaw is an open-source assistant that runs on your own computer.³² Its scheduler, Automations, wakes the agent at set times or on events, such as a trigger from Gmail.³³ It keeps its memory as plain text files on that computer.³⁴ That is the always-on way to continuity in Figure 2.5, on hardware you control.

Models. You choose the model. OpenCode works with many providers, the companies that run models, and with models on your own computer, through Ollama, LM Studio or vLLM. A provider's model runs at that provider, so the conversation goes there.³⁵

Running it safely. OpenCode's agents start out allowed to use every tool. Only reading a file of sensitive settings asks first. So until you set each tool to allow, ask or deny, what it can call is what it may do.³⁶ OpenClaw runs its tools on the computer itself unless you set up a sandbox.³² For an agent a company shares, its security guide says to run it on a dedicated machine.³⁷

The differences that change a decision

Five matter.

1. Where model choice is made. On Anthropic's Pro, Max and Ultra plans,² one model menu sets the model and effort for a conversation that covers both answering and tasks. On Team and Enterprise, chat and Cowork stay separate unless the organization has turned on the one conversation,³ so check the model in each. On OpenAI, an app that still shows Chat and Work has a model picker for each, so the model you tried in chat may not be the one that runs the task. On either AI vendor, record the model and effort for each surface the worker uses.

2. Where shared output is kept. Both AI vendors let a team share a project for standing context. For a document the team edits together, Anthropic offers Claude Docs and OpenAI offers Space. A Claude doc is an artifact, saved with your other artifacts. On Team and Enterprise plans, a Claude doc cannot yet be shared outside the organization.¹⁶ In Space, each colleague brings their own assistant. The AP Worker's register is a spreadsheet, and neither AI vendor's editors include one, so it comes back as a file (Chapter 1). Decide where that file is kept — a shared drive, through a connector — and who can see it, before the first run.

3. Tier names do not match one to one. Compare models by the job each AI vendor gives them, from fastest and cheapest to most capable, and by your own tests, never by name. The same is true of features: the same idea often goes by different names.

   The idea	Anthropic	OpenAI
   A link to another app	Connector	App
   Many sources, checked and cited	Research	Deep research
   A document the team edits together	Claude Docs	Space
   Work that continues between conversations	Scheduled tasks, Claude Duet	A dot
   The AI in a team channel	Claude Tag in Slack	@ChatGPT in Slack and Teams

4. How a standing role gets its continuity. Figure 2.5 showed two ways. Anthropic's scheduled tasks take the second: each run starts fresh and rebuilds from the records. OpenAI's dot takes the first: one agent keeps running, with its own computer. Either can serve a Standing AI Worker, but they put the worker's continuity in different places — so ask the question from Chapter 1: what survives, the machine or the record?

5. Whose name the work runs in. A task that acts as you is fine for personal work. A worker that does a company's work needs an account of its own (2.1). The everyday products work through your own account. Your own dot works on your behalf.²² Claude's scheduled tasks work with your connectors.⁵ In Claude Code, anything a scheduled task, formerly called a routine, does through your connectors appears as you.³⁸ Claude Duet runs on your Claude account, and with your approval it sends email from the accounts you connect.³⁹ Two products can act under a service account: an account that belongs to the team, not to the creator. Claude Tag, in beta on Team and Enterprise, uses its own service accounts in a Slack channel. In a direct message, or with a member's personal connector, the work is recorded under that member's name.⁴⁰ ChatGPT's team tasks, recurring work a team owns on Business and Enterprise, run under the team's service account. Each connection still uses a designated account, with its permissions.⁴¹ A specialist dot gets its own identity, so far only in enterprise pilots.²² So every way to give a worker its own identity needs a team or company plan, or a pilot.

Open source changes who keeps the machine and who supplies the model. OpenClaw takes the always-on way on a computer you control, so what survives is what you keep on that computer and back up. OpenCode lets you choose the provider as well as the model, so write the provider in the Role Contract's runtime needs too.

The invariant

The product names differ, but the way you decide does not. Choose the surface by how much of the work you hand over, the rung on Chapter 1's ladder. For a standing role, get continuity from records or from an always-on agent, not from a machine you hope stays on. Choose the output by what you need to keep afterward. Add each feature, such as a connector or a skill, for the job it does. Choose the model by balancing capability, speed and cost, and test your choice on your own cases. Write the surface and the model in the Role Contract as runtime needs. Then you can change either one without changing the role or what it may do. You still retest the new setup.