Are AI agents here?

The business world is buzzing with talk of AI agents. Google searches for "AI agent" have spiked in recent months alongside a flurry of announcements: Claude computer use, OpenAI's research preview of Operator, and the Perplexity Assistant app for Android.

Search trends for "AI agent" (blue) and "agentic AI" (red), past 12 months. Source: Google Trends

So, are AI agents finally here? And what should we do about it?

In this article, we'll explore:

1. What AI agents are
2. How to evaluate agentic AI
3. So what? Short- and long-term impacts
4. What to do next

What AI Agents Are

The definition of an AI agent is as muddy as the goalposts for AGI, and as confusing as model naming conventions (o1 to o3, anyone? 🤦🏼‍♀️).

There's no clear consensus on what constitutes an AI agent, and some use the term interchangeably with AI, automation, or bots. But to understand what's coming, let's focus on what AI can now do.

Basis defines its agents as "doers not just talkers," which is a definition I like.

To think more broadly about the role of AI in work, I've personally adopted the following framework: "Define, Design, Do." Each task can be broken down into three components:

• Define: Set the goal.
• Design: Plan the steps needed to achieve the goal.
• Do: Execute the plan.

Until recently, most AI was stuck in the "Design" phase. ChatGPT could tell you exactly how to navigate the Domino's website, but it couldn't actually order a pizza for you.

It could walk you step-by-step through how to book a flight online. Tell you the flight time. Suggest the best websites to track flights and prices. Provide the major airlines that travel between destinations. It could even visit other sites and offer details about specific upcoming flights.

But it couldn’t complete the booking. You define, and it could design (which, as we've all experienced, can have real value as well). It still could not do.

However, we're starting to see a shift. Rather than debate whether we’ve crossed the threshold into agentic AI, let’s think of it as a spectrum:

Several recent developments have bridged the gap to true "doer" agents:

• OpenAI's GPT-o1: With enhanced reasoning abilities, o1 can design its own plan to achieve a goal. Even without in-depth instructions or examples, the model is capable of figuring out the best path from point A to point B on its own. It excels in the "design" phase — using a higher level of thinking than what we had seen previously.
• Google's Deep Research: Give Deep Research a research topic (define). It will construct a step-by-step research plan for you to adjust or sign off on (design), and then execute that plan (do), visiting dozens of websites and summarizing findings. While it doesn't take any action in the “real world,” it does arguably execute on a goal by independently completing research-based tasks.
• ChatGPT Tasks & Operator: Tasks allows you to set reminders and recurring tasks within ChatGPT — think AI personal assistant powered by your own prompt library, while Operator can interact with websites to complete real-world actions like making restaurant reservations and placing Instacart orders.

How to Evaluate Agentic AI

When evaluating AI agents, consider these three key areas:

• Scope: What size and scale are the tasks this agent can handle? Current agents and bots primarily excel at narrow tasks ("generate a YouTube video transcript" or "research a sales lead for me"). As agentic AI improves, we'll see broader capabilities emerge.
• Autonomy: How much oversight does the agent need? Does it require detailed instructions? Examples? Reinforcement learning? Or can it learn and adapt independently?
• Access: What level of access does the agent have to the digital world? Is it limited to its own training data (Claude Sonnet)? Can it pull in outside sources (Perplexity Search)? Can it conduct research on other sites (Deep Research)? Can it take action on other websites (OpenAI Operator), or other programs on your computer?

Access in particular has massive implications for privacy and security, and users should proceed with caution. (This awkward intersection of helpful and invasive is perfectly illustrated in Microsoft's controversial Recall feature.)

Beware before giving AI access to things like login details, personally identifiable data, proprietary or confidential information, or your offline computer habits — at least not until you deeply understand how that information is accessed, used, and/or stored.

So What? Short- and Long-Term Effects

The truth is, I believe the impact of this technology will depend heavily on the security and privacy controls offered by the makers of AI agents, and/or the business world's willingness to make these trade-offs (see: The Law of Uneven AI Distribution).

And so in the short term, there may not be massive changes while business leaders seek to answer these important questions.

Long-term, here is one version of what might happen:

Once upon a time (not so long ago), humans carried 100% of their job's workload... All day long, they defined. They designed. They did.

But then people and AI began sharing the load, particularly the "designing" work and "doing" work — at first just a little, and eventually, a lot.

People's workloads got lighter, freeing them up for other things.

This meant:

Increased human capacity. As AI took on more designing and doing, humans had more time to focus on higher-value tasks.

Shift in skills. Demand increased for management skills (delegation, feedback) and strategic thinking.

(Ethan Mollick said of Claude Computer Use: “I was presented finished drafts to comment on, not a process to manage. I simply delegated a complex task and walked away from my computer, checking back later to see what it did (the system is quite slow).”)

Defining work became increasingly important — People's jobs focused less on “How do we get it done?” and instead “What should we do?” Demand increased for critical thinking and long-term planning abilities.

APIs eventually gave way to agent-optimized sites. Some websites, apps, and programs were better optimized for AI agents than others; as a result, humans (or perhaps even AIs) chose these sites, apps, and programs more often.

(See agent.ai creator Dharmesh Shah's take on AI agents and APIs.)

Optimized workflows. Businesses reskilled teams and restructured workflows to leverage people's time and AI's capabilities most effectively.

Statistically average services. With AI self-directing more work, and most businesses using the same handful of leading AI tools, many companies fell into the trap of looking, sounding, and delivering services in very similar ways — the statistical average of messaging, marketing, sales, and services.

Smart companies, however, sought out ways to deliver AI-powered services efficiently while maintaining their unique identity. These businesses had a distinct advantage in differentiation, pricing, brand reputation, and customer loyalty. (More on this in the coming weeks.)

What to Do Next

As agentic capabilities expand, here are some immediate next steps for business owners to think about:

To save money:

• Prioritize operational efficiency. Identify your most time- and labor-intensive tasks, particularly research or administrative tasks with low security risks.

To make money:

• Prioritize scaling services. Focus on low-margin services where AI can improve efficiency or popular services that could be scaled with AI assistance.

In either case:

• Evaluate the risks and benefits of AI agent access, especially regarding sensitive information.
• Consider your preferred level of AI autonomy. Do you want to direct every step, or are you comfortable with AI designing its own path? (From OpenAI: “Users can personalize their workflows in Operator by adding custom instructions, either for all sites or for specific ones, such as setting preferences for airlines on Booking.com.”) This is where rounding up those processes and exemplars will pay dividends.
• Pay attention to the AI agent partner ecosystem. Who and what gets used, and how to make your digital presence more attractive to/compatible with AI agents.  
• Start thinking about how to restructure workflows and upskill your team to focus on higher-value tasks.

Will these AI doers be more disruptive to the workforce than the talkers that preceded them? Probably.

Is this the beginning of the end for human jobs and work? Probably not. (See: Email's story arc from productivity hero to villain.)

After all, to quote Dave Stewart, "The work is never just the work."

‍