The makers' agents, as of November 2026
- Meta's Muse (September 2026): in Meta's words, a personal AI agent that "doesn't just answer questions, it actually does the work", powered by Muse Spark, Meta's own model (Meta).
- OpenAI's Dots (September 2026): always-on agents, powered by OpenAI's own models, that keep working even when the computer is off and can be reached in ChatGPT, Slack and Microsoft Teams (TechCrunch, OpenAI).
- Anthropic's Claude Duet (October 2026): a proactive agent designed around Claude Opus 5.5, in research preview by waitlist, in the United States only (Anthropic).
What they do well
- Integration with the house: Microsoft says Copilot only surfaces company data that each person already has permission to view, and that it honors the sensitivity labels set up in the suite (Microsoft).
- New features first: Muse Spark, Meta's new model, began powering Meta's own products in April 2026, while its API stayed in private preview for select partners (Meta).
- One contract: Microsoft 365 E7 brings the suite, Copilot and Agent 365, which governs agents, together for $99 per user (Microsoft).
These are advantages of design and of contract, with caveats. The bundle does not include everything: Copilot Cowork, Microsoft's agent, requires a Copilot license and is billed separately, by usage (Microsoft). None of the sources for this article includes an independent measurement showing that the maker's agent performs better than a third-party agent using the same model. And early access is narrow: at launch, Muse reached only the United States (and, in October, Singapore), and Claude Duet only the United States (Meta, Anthropic).
First risk: today's best model may not be tomorrow's
On the public human-preference leaderboard of Chatbot Arena, now LMArena, OpenAI's GPT-4 held first place from the leaderboard's launch in May 2023 until March 2024, when Anthropic's Claude 3 Opus overtook it for the first time (Ars Technica). In November 2025, Google's Gemini 3 reached the top (TechCrunch). On November 15, 2026, in the Artificial Analysis Intelligence Index, which combines nine tests of knowledge work, coding and science, and counting each maker's best model, Anthropic was first, with Claude Fable 5.2, followed by Google, OpenAI and Meta (Artificial Analysis). The two rankings measure different things, and neither says on its own which agent serves a company, but both show what matters here: the lead changes hands.
Models retire too. OpenAI promises at least six months' notice for generally available models and three months for specialized variants (OpenAI); Anthropic, at least 60 days (Anthropic). In ChatGPT, OpenAI gave about two weeks' notice that GPT-4o would be retired on February 13, 2026 (TechCrunch). An agent tied to one maker inherits that maker's calendar.
Second risk: for the maker, the agent is one more product
For the company that makes the model, the agent is one bet among many, and bets change priority. The list from recent years:
- OpenAI: ChatGPT plugins lasted about a year, from March 2023 to April 2024 (TechCrunch, Zapier); the Assistants API was shut down in August 2026, twelve months after the notice, and whoever built on it had to rewrite (OpenAI); the end of the Sora video app was announced six months after its launch (CNBC); and the ChatGPT Atlas browser lasted less than a year (TechCrunch).
- Meta: in August 2026, as it focused its AI efforts on experiences powered by Muse Spark, Meta closed AI Studio to creating and editing AI characters; existing ones remained active (Meta). Before that, Workplace, Meta's enterprise network, was discontinued so the company could focus on AI and the metaverse, with two years of transition (TechCrunch); and M, the Messenger assistant, was shut down in January 2018, eleven days after the notice (TechCrunch).
- Microsoft, Google and Amazon: Cortana left Windows in 2023 (TechCrunch); consumer Google+ ended in 2019 (Google), and Stadia in January 2023 (Google); Amazon Chime reached end of support in February 2026 (AWS). These products lasted from three to almost ten years, and still came to an end.
These are examples. Three lists kept by people who track these endings give the count, as of November 15, 2026:
- Killed by Google: 311 products shut down by Google, among them Duplex on the Web, which completed tasks on websites on the user's behalf, ended in 2022, and Gemini CLI, discontinued in 2026 in favor of the Antigravity platform (Killed by Google).
- Killed by Microsoft: 208 Microsoft products, among them Tay, a chatbot shut down in 2016 after 16 hours online, and Teams Live Chat, ended in October 2026 (Killed by Microsoft).
- Killed by AI: 31 AI products shut down by OpenAI, 11 by Anthropic and 7 by Meta, counting retired models, services, apps and acquired startups; Meta's include the Llama API, shut down in July 2026 (OpenAI, Anthropic, Meta).
The risk is rarely a product vanishing overnight. It is that, when the end is announced, the migration falls on whoever depended on the product, sometimes with eleven days' notice, sometimes with two years.
Third risk: disputes between companies reach the customer
- In June 2025, according to Windsurf, Anthropic cut nearly all of its direct access to Claude 3.x models with less than five days' notice; at the time, OpenAI was reported to be acquiring Windsurf (TechCrunch).
- In August 2025, Anthropic revoked OpenAI's own access to Claude models, saying the use violated its terms of service (TechCrunch).
- In August 2026, after SpaceX bought Cursor, OpenAI announced it would end access to its models on Cursor, with a proposed shutoff date of November 12, 2026, saying it could not be confident that SpaceX would use its technology within its terms of service (CNBC).
- Less than a month after Muse launched, Amazon blocked the agent from buying on Amazon.com, alleging a violation of its terms of service (CNBC).
Not every border is a wall: Google's Gemini Enterprise has official connectors for Outlook, Teams, OneDrive and SharePoint (Google), and Microsoft 365 Copilot has one for Google Drive (Microsoft). But what connects, and how far, is each maker's decision, and it can change when interests change.
Five questions to avoid lock-in
- Who owns the history? Ask whether you can export what the agent knows about the company, and in what format.
- How many model makers does it use? And who decides to switch when another model starts to serve better.
- Does it connect to what the company already uses? Calendar, email and documents, whether from Google or Microsoft.
- What happens if the product ends? What notice is promised and what the way out is.
- Is the agent the seller's main product? Or one product among many.
Where Orion fits
The five questions apply to Orion too. The answers:
- Who owns the history? The company. The work records — plans, deliverables, cycles and assessments — belong to it, and canceling the subscription deletes nothing: the organization, the data and the history remain, and Y Managers stays open for reading and editing (plans and leaving). CRM contacts and the deliverables list download as a spreadsheet (.csv) right from the screen.
- How many model makers does it use? More than one: the privacy policy lists the AI providers, among them Anthropic, Google, xAI, OpenAI and Microsoft. Orion Workers decides when to switch, based on the model that serves the customer best today. And a company that prefers can bring its own AI key, choosing the provider and the model (pricing).
- Does it connect to what the company already uses? Yes: calendar, email and documents in Google or Microsoft 365, forwarded emails and a catalog of reviewed tools, such as Canva, Miro and Slack. And in the other direction too: the Claude, ChatGPT or Grok the team already uses can look up what Orion knows about the operation (how to connect).
- What happens if the product ends? The Terms of Service guarantee at least 90 days' notice before a permanent shutdown. During that period, the service keeps running for subscribers, the organization gets, on request, a copy of its data in an open format (CSV or JSON), and whatever was paid in advance for the period that will not be used is refunded.
- Is the agent the seller's main product? Yes. For Orion Workers, the agent is not one product among many: it is the company's focus.
Orion does not depend on one model maker or on one office suite: it is free to connect to everything that matters to its customers, through native integrations, which keep growing — the latest, Instagram, arrived in November 2026 —, and through a connector built on the open MCP standard, through which other assistants query Orion.
What else sets Orion apart:
- A way of working. Orion works with CLEAR, which links the company's objectives to everyday work — key initiatives and KPIs at the top, deliverables plans per team and work plans per person —, and with Y Managers, the system where the team's work is planned, executed, tracked and assessed.
- Personal life and work, kept apart. The same Orion takes care of both at once, and what is personal — family, documents, preferences, dictated notes — stays out of reach of groups and of the company's AI workers (security and privacy).
- Where people already are. Orion talks over WhatsApp, Telegram, the web dashboard and the mobile app, and it is a single conversation: what started in one place continues in the others (channels).
- People to talk to. No one starts alone: there is a first-steps list, a two-minute video, a tour of the screens and a demo company already up and running, and Orion itself guides the setup in the conversation. To talk about the product, support answers by email and on WhatsApp, where Orion replies right away and brings in a person from the team when the matter calls for it — or from the first message, for anyone who prefers to talk to a person.
The plans and the trial period are on the pricing page.
Frequently asked questions
- What is AI vendor lock-in? It is a dependence on a single vendor that makes switching expensive: the history, the integrations and the team's habits stay tied to the seller's product and model.
- Do Muse, Dots and Claude Duet use other makers' models? According to what the makers announced, no: each runs on its own maker's model.
- Do AI makers discontinue products? Yes. ChatGPT plugins lasted about a year, the ChatGPT Atlas browser less than a year, and the Assistants API was shut down in 2026, twelve months after the notice. Meta closed AI Studio to new characters when it focused its AI on Muse Spark. And the lists that track these endings count 311 products shut down by Google and 208 by Microsoft.
- How do you avoid being locked into one AI maker? Choose an agent that uses models from more than one maker, connects to the tools the company already uses and lets you export the history.
- Which models does Orion use? Models from more than one maker; the privacy policy lists the AI providers. Orion switches models when another one starts to serve the customer better.
Sources
- Launches: Meta's announcements of Muse (Sep 8, 2026) and Muse Spark (Apr 8, 2026); TechCrunch on Dots (Sep 29, 2026) and the Dots documentation; Anthropic on Claude Duet (Oct 21, 2026).
- Enterprise agents: Microsoft on Copilot Cowork (Jun 16, 2026), Microsoft 365 E7 (Mar 9, 2026) and Copilot privacy.
- Rankings and model retirement: Ars Technica on Claude 3 on Chatbot Arena (Mar 27, 2024); TechCrunch on Gemini 3 (Nov 18, 2025); the Artificial Analysis Intelligence Index, accessed Nov 15, 2026, and its methodology; the policies of OpenAI and Anthropic; TechCrunch on GPT-4o's retirement (Feb 6, 2026).
- Discontinued products: TechCrunch on ChatGPT plugins (Mar 23, 2023) and Zapier's notice of their end; OpenAI, Assistants API migration guide; CNBC on Sora (Mar 24, 2026); TechCrunch, list of AI projects that didn't make it (Nov 14, 2026); Meta on AI Studio (update of Aug 10, 2026); TechCrunch on Workplace (May 14, 2024), M (Jan 8, 2018) and Cortana (Aug 4, 2023); Google on Google+ (Oct 8, 2018) and Stadia (Sep 29, 2022); AWS on Amazon Chime (Feb 19, 2025).
- Lists of discontinued products, accessed Nov 15, 2026: Killed by Google, Killed by Microsoft and Killed by AI, on the pages for OpenAI, Anthropic and Meta.
- Disputes and connectors: TechCrunch on Windsurf (Jun 3, 2025) and the OpenAI cutoff (Aug 2, 2025); CNBC on Cursor (Aug 29, 2026) and Muse's shopping (Oct 3, 2026); connector documentation for Gemini Enterprise and Microsoft 365 Copilot.
