The Four Horsemen of Agentic AI

Four BIG things in the world of AI this week, from (arguably) the four biggest players in AI.

1. Microsoft Build Announcements
2. Google I/O Announcements
3. OpenAI…a device company?
4. Anthropic Levels Up with Claude 4

Last week was, relatively speaking, quiet. We knew it wouldn’t last.

(logos in the image are respective trademarks of Microsoft, Google, OpenAI, and Anthropic)

Microsoft Build Event

Microsoft’s annual developer conference was loaded with new announcements (and by that I mean a mind-boggling number of announcements; here is the short version, here is the long version) all centered around incorporating AI agents into every aspect of the Microsoft ecosystem. They say the future is an “open agentic web” and it’s clear that Microsoft wants to be in the center of that web. They’re going all-in on agentic AI, providing tools for companies to build, deploy, and manage agents. Here is a select list of some of the announcements (some are live and some are in preview):

• A way to track and organize agents into a library (with Microsoft Entra Agent ID)
• Copilot Studio now supports multiple agents, allowing them to work together
• The ability to tune agents with company data (with Microsoft 365 Copilot Tuning )
• Agents directly in Teams meetings – for you or for anyone in the meeting
• Expanded offerings to give agents the ability to control computers

And many announcements targeted at developers, who will be building and managing vast numbers of custom agents:

• Their coding assistant – GitHub Copilot – gains a coding agent, prompt management, and enterprise controls. Most significantly, agents can work automatically in the background, to keep the coder in the flow of their work.
• An agentic AI platform to design and manage custom AI applications and agents (called Azure AI Foundry) and a similar tool for Windows (Windows AI Foundry).
• An agent orchestration platform called (get ready for this mouthful) Azure AI Foundry Agent Service. It includes all of Microsoft’s agentic tools (such as Semantic Kernel) but is an open platform, with full support for MCP (how agents use tools to accomplish tasks) and A2A (how agents work together).

Google I/O Event

Not to be outdone, Google announced many things at their annual I/O conference this week. Some of the highlights:

• An aggressive expansion of generative AI for internet search, called “AI Mode.” This is a much more AI-immersive experience than their AI summaries, and is a move towards giving you what you need without having to select from a list of links. They really emphasized that it can answer any question no matter how long, complex, or multifaceted. It’s not clear what this will do to their ad revenue, or how search engine optimization (SEO) will be done in this new world (a new term starting to circulate is Generative Engine Optimization, or GEO).
• New pricing tiers (including an Ultra tier at $249/month!)
• A very impressive video generator, called Veo 3 (expensive, part of why they added an Ultra tier). Aside from the usual – better representation of physics (they say “understanding” but it’s still just pattern matching), better item persistence, better character consistency – the huge leap forward is that it generates audio in sync with the video. This is to my knowledge, the first model that does this. There are many examples online, but if you haven’t seen any, watch this fake news. Yes…there’s going to be a lot more fake news coming our way…
• An AI-powered shopping experience to make it easier to buy stuff…or for agents to buy stuff for you
• Deep immersive multimodality in a number of different applications, ranging from letting Gemini see what you see and interact live, the ability to virtually “try on” clothes before you buy them to see what you would look like wearing them, and extended reality (abbreviated XR, this essentially replaces virtual reality and augmented reality) on Android devices
• And a very interesting development: Gemini Diffusion generates text using diffusion instead of transformers. This may be a sign of things to come, let me explain:
  • every LLM currently uses a transformer model
  • every image generation app currently uses a diffusion model
  • a transformer model generates one token at a time
  • a diffusion model tries to create all of them at once
  • Therefore, diffusion is faster. If you’re even slightly interested in this, I recommend you watch their 90 second illustration here. 

But the stat that amazed me was that token consumption has increased 50x over the past year (token consumption is a good measure of AI use). Some of that is growth in the number of users, some of that is because people are using more AI, and some of that is that models are using more “reasoning” – in reasoning, the models “think” through their answer before composing a final answer, and all this “thinking” is done by generating more tokens. 50x is a big increase, and we can expect it to continue to climb quickly.

Speaking of more reasoning, Google added a “Deep Think” mode for Gemini 2.5 Pro that spends more time “thinking” to provide better responses. This reinforces the trend of applying more test-time compute to scale the capabilities of these models.

OpenAI Enters the Device Market

OpenAI acquired (subject to approval) Jony Ive’s one-year old startup, io. For a whopping $6.5B! Jony Ive is famous for designing the iPhone. His company, io, was working on the “next thing” in devices. With this acquisition, OpenAI plans to create a family of AI-powered devices. Jony Ive and his team will lead design and creative at OpenAI, and io will merge with OpenAI. Sam Altman says to expect a product next year, and (he never misses the opportunity to go over-the-top) teased that it will likely be the most amazing and revolutionary device ever.

Anthropic Levels Up with Claude 4

The Claude 4 model family is out, with Claude Opus 4 (the big model) and Claude Sonnet 4 replacing the 3.7 series (no word yet about the smallest model – Haiku). Apparently, they’re really good. Of course, they improve on the usual benchmarks, but it takes time to truly evaluate how much those improvements translate to real applications. But early feedback is very impressive and they’re touting some major advancements:

• Claude 4 is now the best model for coding. In one example, Claude was given a coding task and Claude worked on it for seven hours straight – while duration is a weak proxy for quality, that’s an order of magnitude above any previous model.
• Claude 4 can use tools in extended thinking mode (what they call their “reasoning” mode), not just the regular, quick response mode. Tool use is critical for expanding capabilities, so incorporating tools in this way is a huge boost to what they can accomplish.
• These models can now use tools in parallel, which provides faster responses especially for complex tasks, and have better memory use (it will choose important facts and automatically store them in a local file if it’s given permission) to keep track of important things in longer, multi-step processes.
• Most significantly Anthropic is positioning Claude for as a hybrid reasoning agent that helps you get work done.

So, the one-upmanship of each of the big model houses continues (although right now it seems the arrows should be pointing in the other direction):

(image courtesy of @piet_dev on X)

However, the “which model is the best” is becoming less and less important. It’s going to be less about the model(s)…it’s going to be a lot more about how those models are used, and how they’re incorporated into our lives at work and at home. 

---

---

All Agents, All the Time

We knew 2025 would be the year of the agent, but the pace of this change seems to be faster than most expected. Everyone is pushing hard into agentic AI. Microsoft’s event was all about agents – standard agents, custom agents, and agentic AI: how to build, manage, and orchestrate multiple agents. Goole’s event was similar, but less focused on business and more focused on consumer. Anthropic is changing Claude’s positioning, from a conversational language model to a conversational agent.

We’re catching a glimpse of the future of agents for work, agents for home, agents for life, agents everywhere.

Agents for Work

Microsoft is imagining agents everywhere at work – agents that help you, agents that work alongside you, agents that do some of the work instead of you. A multiagent world where agents collaborate with one another. They want to be at the center of this ecosystem, so that whether you use agents they provide, agents built within your company, or agents you create, it’s all Microsoft.

There is value in having a single platform to manage all of the agents in an organization – that gives you the best control to govern, monitor, and manage all the agents. But if every major player is pushing an agentic ecosystem (they’re now available in various forms from Microsoft, Google, IBM, Amazon, Salesforce, Sitecore, Dataiku, and many others) it’s unclear if companies will be able to select and use a single platform, or if a multi-platform model will prevail. Large companies WANT to standardize on one platform, but often it’s impractical. Take cloud for instance: most large organizations have instances of AWS, Azure, and GCP. Sure, one dominates…but all are present.

The future of this is uncertain, so companies are racing to be THE hub for agentic AI in companies.

Agents for Home

Google is incorporating more AI into everything they provide. Of course an AI-powered search experience, but into your everyday life with extended reality on Android phones. Instant information access. Shopping. Organization. Communication. Help with everyday tasks. An always-on AI agent, ever ready to perceive what you’re experiencing and step in to help. Remember Google Glass? Now the technology is there to make it happen, and although they haven’t teased bringing back glasses (unlike Meta which already has AI-powered Ray Bans and Apple which apparently has glasses planned for 2026) you have to believe that they’re evaluating that in addition to Android phones.

I’ve seen Google criticized this past week for being so consumer-oriented and not targeting businesses. I agree that their I/O conference was consumer focused, but they are definitely targeting business – they already have Google Agentspace which has (or plans to have) all the functionality that Microsoft rolled out with their agentic AI announcements. So I see the announcements focused on the respective sweet spots of these two companies, not an indication that Google isn’t interested in or doesn’t have a strategy for businesses. In my view, Google’s technology is more advanced than Microsoft’s (which doesn’t even have their own models – they’re still tied closely with OpenAI). Microsoft is still struggling with consumer (when was the last time you used Bing?) and Google doesn’t have the presence in enterprises that Microsoft does. So they’re capitalizing their respective strengths first, and only secondarily trying to move into the other one’s space.

Agents for Life

OpenAI is betting that the future of AI is devices. It’s not their only bet, but it’s a very big one ($6.5B is their biggest acquisition by far, and they’re still bleeding cash). A device that you carry around all the time (like your phone, but probably without a screen) that sees everything you see and hears everything you hear (and say). A device that serves as your companion, colleague, and friend, with a perfect memory of everything you’ve ever experienced to seamlessly helps you through your day.

Or something like that. I guess we’ll see next year. The AI device approach was tried before, and the tech wasn’t ready (the Humane AI pin is defunct, and the Rabbit R1 is not getting much attention). The tech is already much better, and by next year it will be leaps ahead – so that won’t be the constraint. The application of the technology will be the make-or-break thing here: will it be so useful that everyone will overlook the privacy and surveillance concerns of letting an AI experience everything that you see and do?

Agents Everywhere

Anthropic doesn’t have the company customer base that Microsoft has or the consumer base that Google has, so they’re focusing on making what they do have – Claude – agentic. What before was a chatbot or conversation partner, or maybe a conversational search utility, is now an agent – something that figures out what you want, the best way to get what you need, and then does it. Their initial strength is with code, with Claude 4 now being the best available model for coding.

Much of this is done with tools, which are other software programs that the agent can use. It’s no longer limited to working only with conversation; for example, it can use a calculator tool to do math or form tool to submit a warranty request or a reservation tool to book an airline flight. The number of tools available right now is still small but that will not last long, which will rapidly accelerate the capabilities of agents.

The Takeaways

What we see here is an incredible push to agentic AI. An expansion of AI capabilities as agents start to not just use tools, but use multiple tools and use them in parallel. An incorporation of AI into every product and every experience. AI is coming for every part of our lives, whether we want it or not. It’s not even clear how much we’ll be able to resist, since it’s coming to all of the tools that we already use, and it promises gains that we won’t be able to easily ignore.

Friction is the problem. We are in the stage of making AI easier.
They’re not adding features. They’re removing steps.
– Conor Grennan, the AI Mindset

With this past week we have crossed a threshold. While companies like Anthropic are still releasing better and more capable models, the story is shifting away from models. It’s shifting to agents – to the application of models that actually do useful things. And to abstracting away the models so that you don’t have to work with a model, but the AI is embedded into things that you use and do every day. The AI is starting to disappear behind the scenes so that it just works, giving you a better experience and better results. This was the vision behind Apple Intelligence – it’s suppose to “just work” in the flow of things you would already be doing anyway. Now everyone is chasing that goal: make the AI seamless, invisible, and, in the words of Conor Grennan, frictionless.