With the increasing number of AI models, does a company have to rebuild an entire system every time it switches models?

If AI can already help with writing articles, creating presentations, and analyzing spreadsheets, why do we still keep copying and pasting between different apps?

And what does it mean if OpenAI, ChatGPT’s developer, might buy AI chips directly from its competitor Google?

Today, October 29, 2026, these three key announcements highlight the changes happening in the AI industry.

The competition is no longer just about which model is strongest, but who can reliably complete tasks while controlling software and compute costs.

1. Microsoft Opens Copilot Harness: Why Companies Don’t Want to Rewrite Workflows for Every AI Model

On October 28, Microsoft announced the public preview of Copilot Harness within its Microsoft Foundry Agent Service.

This release is not another chatbot, nor a new large language model.

Instead, it addresses fundamental challenges AI agents face in real-world workflows.

Imagine a company has built an agent that daily needs to:

  • Read internal documents.
  • Query specific data sources.
  • Analyze data.
  • Run programs.
  • Generate reports.
  • Finally, hand off parts that need human approval to managers.

This work isn’t just about passing questions to an AI model for answers.

Someone has to manage tool calls, execution environment, data context, permissions, and task progress.

Microsoft’s goal is to provide this integrated execution framework.

What exactly is Copilot Harness?

You can think of the AI model as the part responsible for thinking.

But to complete real tasks, a full mechanism is needed to decide what tools it can use, what data it can access, and when to pause the process.

This mechanism is called Harness.

Microsoft says the new Copilot Harness is based on a multi-model execution framework developed by the GitHub team and already used in various Microsoft and GitHub products.

The framework supports over 20 leading models and can accept user-provided model keys.

However, this is the theoretical capability; the public preview in Foundry limits available models and tools based on support lists, regions, and preview status.

The key is not that each agent uses 20 models at once, but that developers don’t have to lock a workflow to a single model.

What features are open now?

Microsoft divides the public preview functionality into four main parts.

First, reusable Skills.

Developers can package fixed work procedures, programs, and resources into Skills.

For example, creating a reusable research routine that reads reports, verifies sources, summarizes content, and flags gaps.

Second, connecting external tools.

Agents can use supported Foundry tools and Model Context Protocol (MCP) tools with configured authentication and approval mechanisms.

Third, execution and task delegation.

A main agent processes documents and programs in a controlled environment, delegating limited work scopes to sub-agents and consolidating results.

Fourth, workflow version control.

Developers can manage agent definitions and resources alongside their software projects for easier testing, tracking, and modification.

The shared goal is to let enterprises focus on business rules and result quality rather than rebuilding the agent infrastructure each time.

Can AI pick up automatically if interrupted mid-task?

Microsoft also announced that Foundry Hosted Agents now officially support long-running task recovery.

This feature allows restarting a task after an interruption.

However, an often overlooked limitation is that:

The platform can restart a task, but the application won’t automatically know which steps are already done.

Developers still need to save progress and intermediate results themselves.

For example, if an agent is creating a report and the task is interrupted, restarting without checking if the report already exists may cause duplicate documents.

Likewise, if tasks include updating databases, sending notifications, or creating orders, duplicate execution must be avoided.

Thus, "long task recovery" lays the foundation for recoverable workflows, not that AI will never err.

Is this feature available to general Microsoft Copilot users?

No.

The public preview targets developers and business teams building agents in Microsoft Foundry.

It doesn’t mean every user of Microsoft Copilot or Microsoft 365 suddenly has a Harness button available.

Moreover, Copilot Harness and Hosted Agents’ long-running recovery are at different release levels:

  • Copilot Harness is in public preview.
  • Recovery capabilities for Hosted Agents are fully released.

Is there a fee to use Copilot Harness?

Microsoft states that starting December 1, 2026, Copilot Harness will charge usage fees for execution environment hosting.

Before then, Harness Runtime use is free.

However, fees for model inference, individual tools, and other connected services may still apply.

Microsoft’s announced reference pricing as of December 1 is:

  • $0.0994 per vCPU hour.
  • $0.0118 per GiB memory hour.

Prices may vary by region and contract.

This also reminds companies: using a common agent platform still requires evaluating total task cost—not just comparing model token prices.

What’s the biggest significance for enterprises?

AI systems previously might be built around a single model.

Now companies increasingly need to consider model quality, speed, cost, and tool support.

If every model switch means rewriting entire programs, costs keep rising.

Microsoft aims to keep models and external workflows as separate as possible.

But switching models doesn’t guarantee identical results.

Model tool usage, prompt compliance, and answer quality still differ.

Even with a common Harness, companies should retest before formally switching models.

The truly important point:

Models can be interchangeable, but work rules, permissions, and acceptance criteria can’t be lost when changing models.

2. Apple Adds AI to Pages, Numbers & Keynote: Mac Users May Finally Avoid Constant App Switching

Another update closer to general users comes from Apple.

On October 27, Apple announced new features for Apple Creator Studio.

Pages, Numbers, and Keynote will each gain AI assistants that understand the current document content.

For the first time, Apple is adding MCP support to these office apps, enabling external AI agents to interact with the content, perform edits, and automate tasks through a compatible interface.

What can Pages, Numbers, and Keynote do?

Pages: Assist with writing and editing documents.

For example, users with a report can ask AI to adjust structure, revise paragraphs, or organize drafts.

Numbers: Help analyze and organize data.

Besides calculations, AI can recommend suitable chart types with one-click previews and assist in creating custom tables.

Keynote: Aid in creating presentations.

AI supports slide design, adding animations, and adjusting content.

The common approach is AI directly understanding the file being edited, not forcing users to copy contents into another chat window every time.

However, Apple’s announcement describes future directions, not guaranteed correctness in every scenario.

Facts in documents, spreadsheet formulas, and numbers in presentations still require human verification.

What is MCP? Why is it more noteworthy than just a built-in AI assistant?

MCP stands for Model Context Protocol.

Simply put, it’s a standard protocol that lets AI apps exchange information with external tools and perform supported operations.

With Pages, Numbers, and Keynote supporting MCP, compatible external agents could not only offer suggestions in chat but directly interact with the document through tool interfaces.

For example, an authorized workflow might:

  • Read existing reports.
  • Organize key data.
  • Assist updates to spreadsheets.
  • Prepare presentation drafts.
  • Then hand off to users for review.

This is a possible workflow design, not an Apple guarantee that all third-party agents can freely perform every step.

What can be done depends on compatibility, permissions, and actual product capabilities.

Will Taiwanese users get this on October 29?

Not yet, and don’t consider it fully available.

Apple officially states the Apple Creator Studio update is planned for November 2026.

The AI assistants in Pages, Numbers, and Keynote are currently in beta.

There are three important restrictions:

First, language limitation.

The beta assistant initially supports only US English.

So, full support for Traditional Chinese documents can’t be claimed yet.

Second, device and OS requirements.

The AI assistant requires Apple Intelligence-supported devices running iOS 27, iPadOS 27, or macOS 27 or later.

Third, MCP has separate constraints.

Pages, Numbers, and Keynote’s MCP features require macOS 27 or higher.

So don’t confuse the AI assistant and MCP device requirements.

Is a subscription required for this update?

Apple Taiwan announced Apple Creator Studio new subscriber pricing as follows:

Plan|Price (TWD)
General Monthly Subscription|NT$390
General Annual Subscription|NT$3,900
Eligible Education Monthly Subscription|NT$90
Eligible Education Annual Subscription|NT$900

Existing Apple Creator Studio subscribers will receive this update in November without extra purchase.

But this doesn’t mean all free users of Pages, Numbers, or Keynote automatically get full AI features.

Apple confirmed free versions of these apps will remain available.

So don’t rush to pay for AI features just because of the news. Wait for full launch, then decide based on language support, device compatibility, and actual functions.

What does this mean for everyday workers?

It reflects a shift in office software.

AI used to be a separate tool.

Users prepared documents, gave data to AI, then pasted results back into apps.

Now, Microsoft, Google, and Apple are moving toward AI directly working inside familiar apps.

This saves repetitive copy-pasting.

But it raises the challenge of version control, editing permissions, and undoing changes.

For official quotes, financial data, or externally sent documents, the safest approach remains generating drafts, preserving originals, and confirming manually.

3. OpenAI Reportedly Planning to Buy Google TPU: Why Would a Competitor Buy Chips from Google?

The third story comes from Reuters on October 28.

Reuters, citing The Information and an insider source, reports OpenAI is planning to purchase AI chips from Google.

There is no public confirmation of contract size, chip quantity, or delivery terms.

Google did not immediately respond to Reuters, and OpenAI declined comment.

So the accurate statement now is:

OpenAI reportedly plans to buy Google TPUs, but no official signed large-scale purchase agreement exists yet.

Doesn’t OpenAI mainly use Nvidia chips?

Yes.

OpenAI has long relied heavily on Nvidia GPUs for AI training and inference.

But the growing compute demand makes relying on a single supplier risky for cost, supply, and deployment flexibility.

Google TPU is a different AI accelerator, not a direct Nvidia GPU replacement.

Different models, architectures, and workloads perform differently across hardware.

Using multiple chip types doesn’t guarantee every AI task becomes cheaper immediately.

Consider also the cost of porting models, adjusting software, performance testing, and infrastructure management.

Did OpenAI use Google TPU before?

According to Reuters, OpenAI started renting Google AI chips in 2025.

The difference now is the potential move from renting compute to directly buying chips.

No details yet on the purchase scale.

Therefore, we can’t infer OpenAI is abandoning Nvidia or that Google TPU is their primary compute source.

OpenAI also collaborates with Broadcom to design proprietary AI chips.

This signals large AI companies exploring diverse hardware options.

Why would Google sell chips to OpenAI?

OpenAI and Google compete in AI models, search, and some products.

But Google also runs a cloud platform and AI infrastructure business.

If OpenAI becomes a customer for Google TPU, the two companies can compete on products but cooperate on infrastructure.

This dual relationship is common in tech:

One company can be another's rival as well as supplier or customer.

Google is expanding TPU external sales.

For example, Oracle announced on October 27 plans to bring Google TPU to Oracle Cloud Infrastructure in 2027.

Oracle’s agreement is an official future plan.

OpenAI’s TPU purchase remains a media-reported plan.

They differ in confirmation level and shouldn’t be confused.

What does this mean for Nvidia?

If major AI developers increase use of alternative compute architectures, Nvidia may face more competition long term.

But current information is insufficient to predict Nvidia’s revenue impact based on this report alone.

Key factors to watch include purchase volumes, delivery timing, and actual AI workloads handled by the new chips.

OpenAI still needs extensive training and inference resources.

Adding another supplier doesn’t automatically mean less Nvidia spending.

The main industry takeaway is:

Large AI companies now compete not only on model quality but also pursue diverse, cost-controllable compute sources.

4. What Is the Common Meaning of Today’s Three Announcements?

Microsoft wants companies to use AI agents without rebuilding full workflows for every new model.

Apple brings AI assistants and MCP into daily documents, spreadsheets, and presentations.

OpenAI reportedly seeks more AI chip procurement options.

On the surface, these address three different markets.

But they each tackle a different AI system layer.

First Layer: How models actually execute work.

It’s not just the answering ability—it requires tools, execution environments, permissions, and task recovery.

Second Layer: Can AI work directly inside software users know?

Reducing copy-pasting can improve efficiency but demands clear edit controls and checks.

Third Layer: Who provides compute resources that back all work?

As AI services spread, chips, power, vendor choices, and infrastructure costs become more critical.

All layers are essential.

A smart model alone cannot reliably complete real tasks without tools.

An AI agent controlling many apps without clear permissions can cause risks.

And a popular product with uncontrolled compute costs risks sustainability.

5. What Should Regular Users Do Today?

If you use ChatGPT, Copilot, or other AI tools daily, there’s no need to switch services immediately based on today’s news.

Microsoft Copilot Harness currently targets developers and enterprise agent builders.

Apple’s new office AI features launch in November with language, device, and subscription limitations.

OpenAI’s Google TPU purchase still lacks official transaction details.

Instead, develop a habit of asking three questions when evaluating AI products:

  • Can it not only answer but reliably complete tasks?
  • What data and software can it operate on, and are permissions and confirmation clear?
  • Is the claimed cost advantage real or still speculative?

Fast AI development means distinguishing official announcements, test features, future plans, and media reports is crucial.

Because what truly impacts work and life isn’t the capabilities companies announce, but whether those capabilities can be used safely and reliably.

The Most Important Sentence Today

The next wave of AI competition isn’t just about the smartest models but about who can integrate models, tools, and compute into truly reliable work systems.

Recommended Reading