Weekly Links: A Deep Dive Into "Over the Top" AI Agents in the Enterprise

This week, we had an amazing "3claw summit" organized with some of my previous colleagues from 3scale. So great to hear what everyone is up to (including with AI!) - on the cutting edge in many cases! In the wider world, we got new OpenAI GPT 5.4 Nano and Mini models, human brain cells are playing Doom in Australia, and does it make sense to blame agents for leaks?

For the main part of the post this week, I'll change the format a little and dive into just one topic, with several associated links. The topic seems important enough to warrant deeper analysis, and so here we go. Let me know what you think of the format!

I'm also posting this on Commune, so join me there for discussion:

"Over-to-Top" in the Enterprise

In the past couple of weeks, a few things happened that might seem disconnected but aren't. Together, I think they represent a phase change in how AI is being deployed and in the market structure that will result. The change may reverse, but if it sticks, it's a big deal. Specifically, this week, OpenAI announced it would cut back on AI and refocus on productivity (why would they do that?). There has also been an enormous amount of turmoil in the public markets for software companies as AI and Anthropic's Claude gains traction in the Enterprise. Lastly, there was the announcement by Microsoft of its new Anthropic-powered Microsoft 365 CoWork product, which I've already commented on previously.

In one way, these are obviously related: AI Agents are gaining traction in the Enterprise, and both incumbent software vendors and OpenAI are starting to see Anthropic taking a lead. But what is more important is how they are diffusing.

• Up until a few months ago, there has been an obvious path for AI into the enterprise: being infused into existing software layers from "underneath". In other words, OpenAI, Anthropic, and other providers being a new kind of infrastructure that lends Intelligence to all the layers above. This was provoking a race amongst software providers and enterprises themselves to embed more AI and layer on top of the AI APIs they could call. A case in point was Microsoft's rollout of MS Copilot across all it's windows surfaces. There were places in which the frontier model labs were competing directly for business (e.g., with the Claude for Finance offering), but this was relatively rare.
• Now things have changed. In the last 3-4 months, both Anthropic and OpenAI have increasingly promoted Agent harnesses that use their latest models and provide a user experience that is well beyond just "model use" via API. Specifically, with the harness and the model together, it's possible to tackle much bigger tasks that could run for hours before concluding. This works especially well in coding, where agents can check their own work and just keep going if there are errors. However, it's also useful in general knowledge work. Claude CoWork, in particular, is good at extracting information from documents, entering it into a spreadsheet to build a model, and then using the results to drive a web form. This all requires access, but the use cases are increasingly compelling.
• The fact that Microsoft has moved away from its previously trumpeted model-agnostic strategy (ingesting intelligence from the bottom up and making the model subordinate) to embracing Claude CoWork as the brand for its new offer and committing to one model is eye-opening. This is a huge strategic win for Anthropic since, for this product, CoWork is now "over the top" of the entire Microsoft Office software estate. It's also an admission from Microsoft that they were unable to build anything as good for an agent harness as Anthropic's harness + model combined. Ben Thompson, writing on his excellent Stratechery blog, puts it something like this: the coupling between harness and model is clearly so strong at the moment that no third party can replace it. I suspect that this might be reversed and decoupled again, but for now, with their super agents, the frontier model labs have the upper hand.
• OpenAI is a little less visible in this regard, but it does have its own Enterprise Agent offering called Frontier, and it also recently acquired the Agent testing and prompt management company Promptfoo. The company's Codex AI is also a very good alternative to Claude Code, which is the coding agent from Anthropic that Claude CoWork is based on.

So why is this all a big deal? It looks to me that OpenAI's "refocusing" is really a very clear recognition that the stakes in Enterprise just got way higher. When AI was diffusing "from below" through existing software layers, work was done by partners. The model providers would likely split the pie on who powered whichever software layers won. They'd also try to take the most valuable pieces for themselves.

If, however, a class of Super Agents that sit on top of all other software and orchestrate it, this is a race for supremacy. It's also why SaaS stocks have wobbled even more. Many of the players in the SaaS space don't have an answer to just being a database that an OpenAI or Anthropic agent manipulates. OpenAI's urgency likely also comes from an understanding of how powerful owning the entry point to productivity is. This is exactly what has happened in the consumer space: ChatGPT has become the default entry point for AI experiences. The idea that Anthropic could become the same for the Enterprise is a real emergency.

In the long run, I expect other players will also be able to hold their ground, but this is a big breakout moment.

I'm using the word "super agent" here to refer to those "general agents" like Claude CoWork, OpenClaw, and others that are effectively orchestrators of other tasks. This is similar to the notion of a "super app," which we saw emerge on mobile. This also distinguishes these systems from more standard agents that perform specific tasks in an automated way.

An additional piece of the puzzle can be found in a less widely diffused but great interview by Felix Reisenberg, who leads Claude CoWork at Anthropic. He joined the Latent Space podcast this week (in a very techy session, but I think it contains a lot of insights), and the key thread in his message was: give your (super)agents their own virtual machine. This points to a future where Anthropic may be betting on hosted "self-contained" systems that manage user context and permissions, whereas OpenAI has likely been operating under the assumption that people's personal agents would run multi-tenant on their own servers.

This is also what made Microsoft's CoWork announcement quite unique. Using that product means using CoWork running within Microsoft's cloud but tied into your Enterprise IT environment. Arguable a half way house between local deployment and multi-tenanted on Anthropic's own servers.

This "over-the-top" change is really an important potential shift, and one might ask if existing software vendors can fight back. They likely can, in particular because adopting such a high-trust system will take time in many places, but it will be hard. The "productivity super agent" has some key advantages in the long run:

1. They can access (if you let them - which is still an open question), the interfaces, CLIs, and APIs of many different applications, so they are able to fuse workflows across tools.
2. They are a good solution for management of what data can be shared, what tasks can be delegated etc. on a person by person basis.
3. Having a single interface/control point for all productivity is a really powerful hook that will be hard for people to move away from once it takes hold.
4. Down the line, it will be logical for these personal super-agents to interact with those of other people (and with shared work products), which creates a strong potential network effect certainly within a single organization, but potentially across organizations. This may end up leading to standard protocols (and hence less vendor lock-in), but using the same system will likely be lower friction for quite some time.

Perhaps one last consideration is interfaces and workflows.

Interfaces

A chat window is not the right interface for many work tasks, and so won't existing software persist for that reason?

For a while, yes. "Over to top" means agents are using those same interfaces humans use. But over time, they'll use APIs and CLIs, perhaps even switching out the apps. Interfaces will also become much more custom and fluid.

Into this bucket I'd also throw Google's update of Stitch, which is now a fully fledged UI design system and can not just produce beautiful designs but deliver the underlying web/UI components that can later go into code. This is another step closer to "completely dynamic" UI, which just gets created and rendered on demand rather than being pre-designed.

Workflows

Perhaps the stickiest things in Enterprise are workflows: processes that involve multiple humans and hence guide the "doing of business" in an organization. Often workflows also involve physical systems, sensors, customers, points of sale, and many other complex assets.

Also here, the assumed mode of AI diffusion has been "bottom up" with vendors who built these workflows integrating AI.

It is very likely that many of the workflows and the software that powers them will stay around for a very long time. Ripping such things out is just too painful. Over time, however, general competence AI layers on top (and in particular super agents) will become the operators of the workflows and likely the ones adjusting and maintaining the workflows.

Vendors here won't be replaced, but they might lose valuable contact with their end customers and become infrastructure.

What's next?

It seems that OpenAI has realized that it cannot afford for Anthropic to get too far ahead on Enterprise revenue and will fight hard to keep parity. Existing non-AI players will have to figure out which parts of their AI offering will stay in the customers' hands and which might start being activated by agents that come on top of the line.

Still, other players will be trying to figure out whether they can build Agent harnesses that are model agnostic and do better than those tied to a single vendor.

All change in enterprise software (again!).