For the past eighteen months, Silicon Valley has been buzzing with promises of AI agents — autonomous AI assistants that can act on our behalf, handling everything from scheduling meetings to writing code for applications.

OpenAI’s Sam Altman has been particularly vocal, predicting since mid-2023 that AI agents would enter the workforce and replace entire job categories by 2025. (Note: He’s also said AGI is almost here … and then told people to stop hyping it up.)

In 2024 we got a glimpse of the agentic future – with experimental releases from Devin (an AI coding agent), Project Mariner (a browser-based action-taking agent), and Anthropic’s computer use agent.

But the 2024 agents weren’t quite there yet. Some were impressive first tries, but most underwhelmed us when we tested them in the real world. Specifically, the agents lacked the ability to maintain context across actions (like fixing a bug in one code file while breaking three others) and didn’t have the proper tools to execute tasks (like getting stuck when trying to book a flight because it couldn’t handle the payment form).

2025 has introduced something more interesting: two distinct philosophies on AI agents, via the release of OpenAI's Operator and Microsoft's Copilot Studio.

Big Tech’s very different approaches to agents

Microsoft and OpenAI have released, in our opinion, two “takes” on agents:

• The bounded agent (Microsoft Copilot Studio)
• The open agent (OpenAI Operator)

With the bounded agent, Microsoft is betting that enterprises want predictability over possibility – systems that work within defined parameters and never color outside the lines.

Copilot Studio is like a well-trained, but junior, office worker: it knows exactly what it can and can't do, and sticks to the script.

It can compile team progress reports, draft meeting agendas based on project priorities, and search internal knowledge bases for solutions – but only with carefully defined guardrails.

It can even handle employee IT issues by following pre-approved playbooks in SharePoint, but it won’t venture beyond the prescribed boundaries or attempt creative solutions outside its instructions.

On the other side, we have OpenAI’s Operator – an open agent.

Operator is messier and less predictable, but potentially far more powerful.

Unlike Microsoft’s bounded approach, Operator can interact with any website or interface it encounters, adapting to the digital world as it goes.

It can book your flights, order groceries, and handle tasks in your browser without needing pre-programmed instructions. But this flexibility comes at a price. It might click the wrong button, misinterpret form fields, or get lost in navigation menus. It’s like having an eager intern who can figure out any task – but only after doing it wrong a few times.

Which is better? Maybe neither.

The power of Operator lies in its ability to tackle unexpected challenges without instructions— something bounded agents simply can’t do.

But let's be honest: both approaches are still finding their feet.

When using Operator to grocery shop, it took me 15 minutes to add chips to my Instacart cart – something I could've done in 30 seconds. It couldn’t even find the specific brand of chips I wanted. It was like watching someone try to navigate a city with a map held upside down.

And Microsoft's bounded approach means their agents often hit invisible walls, unable to complete tasks that require even slight deviation from their instructions.

Here's a real example: I built a simple agent to recommend Section’s courses based on a person's role and learning goals. Despite clear instructions to only suggest Section's courses, Copilot’s agent immediately started recommending Coursera and LinkedIn Learning content instead. Ouch.

I had to explicitly restrict its web access and limit it to only searching our website — exactly the kind of rigid boundaries that make bounded agents both reliable, but so frustratingly limited.

So… is the hype real?

The limitations of both approaches might seem like deal-breakers, but we're witnessing the first stumbling steps toward agents that can truly act on our behalf. This isn’t the “AI will replace all jobs” hype we’ve heard before. It’s about creating systems that can collaborate with humans, understanding both the explicit rules and implicit norms of how work gets done.

Microsoft’s bounded agents show us a future where AI can independently navigate between enterprise systems to manage workflows. OpenAI’s Operator demonstrates AI’s future ability to adapt to new digital tasks in real-time.

What we get to spectate in 2025 is how these two fundamentally different visions of AI agents will shake out. The bounded approach promises safety and reliability that many enterprises need for their first steps into AI agents. But these training wheels may prove too restrictive as organizations mature in AI.

Just as many companies initially adopted locked-down enterprise software before embracing more flexible cloud solutions, we might see a similar evolution here: Starting with bounded agents for basic tasks before graduating to open agents like Operator that can handle more complex and dynamic challenges.

My verdict: If you're knee-deep in the Microsoft ecosystem, don't watch from the sidelines – take 10% of your workforce and give them access to agents in Copilot Studio. Yes, you're partly paying to be a beta tester, but you're also buying your company a head start.

For the rest of us? Play with Operator. Let it fumble through ordering your groceries. It's messy, it's imperfect, but it's a front-row seat to the future.

Here’s our predictions for how the war of the agents will play out in 2025:

1. Microsoft will continue to dominate enterprise, but stumble with innovation. Their approach will win over cautious IT departments and pass infosec reviews, but they’ll struggle to evolve past their current boundaries.
   
2. A hybrid approach will emerge. We’ll see the first enterprise tools combining the best of both approaches. Bounded agents handling sensitive internal systems (HR portals, ERP systems), while open agents tackle external facing tasks like customer support and market research. Your supply chain agent gets full clearance, while your LinkedIn outreach bot gets basic web access.
   
3. Operator will have an “iPhone moment”: Early adopters will become Operator’s training ground, helping it learn to browse the web and perform tasks. And for most people, their first real experience with an open agent won’t come from startups or competitors — it’ll be through OpenAI.‍
4. The “open” race will heat up: Google and Anthropic won't let OpenAI own the consumer-friendly agent space. While both already have capable agents (Project Mariner and Claude's Computer Use), they're still too technical for mainstream use. Expect Google to transform Mariner into a Chrome-native experience that anyone can use, while Anthropic repackages Claude's capabilities into a more approachable and "responsible" agent. By year's end, we'll have three competing visions of how open agents should work with humans, each racing to be as accessible as Operator.