Your next coworker could be a machine. Think about that for a minute.

When asked about his plans to expand Nvidia, CEO Jensen Huang replied, “Nvidia has 32,000 employees today. I’m hoping that Nvidia someday will be a 50,000-employee company with 100 million, you know, AI assistants in every single group.”

For much of 2024, the buzz was about so-called ‘agentic AI,’ the field of AI in which models have agency, which means they can use tools to complete tasks. Agents will come of age in 2025, and deployments will soon begin at scale. Many have declared this year to be the year of agents. In this post, I’ll review why that’s the case, what types of agents are coming, and the implications for the future of work in 2025 and beyond.

Reasoning is the foundation of agentic AI

We often judge the intelligence of other species by their ability to choose tools and use them to perform tasks. Chimps use sticks to dig juicy termites out of their mounds, elephants can paint pictures, and crows curve sticks into hooks to probe for food hidden in trees. Intelligent AI agents need to choose and use the right tools to perform steps in complex tasks. They also need to be able to plan how to execute those steps and perform them in the correct order. This requires reasoning abilities, or they will quickly fail.

To complete complex tasks, AI agents may need to accurately use tools 20 to 100 times or more in a sequence of events. If they choose the wrong tool 10% of the time or misuse it (e.g., clicking the wrong button on the screen when controlling actions through a browser) over 100 steps, their chances of correctly completing the task are vanishingly small (0.9^100 = 0.002%). This is where reasoning comes in. AIs with improved reasoning capabilities might use the right tool the right way 99.95% of the time, dramatically improving their chance of success. Think of agentic AI as a simple stack of capabilities: agents that run on top of reasoning technology that runs on top of large language models.

Reasoning capabilities have received a nice boost recently. OpenAI’s o1 and o3 models use chain-of-thought reasoning to develop multiple hypotheses for solving a problem, assess the best path forward, reflect on their approach, and correct their mistakes. This approach has led to impressive leaps in reasoning abilities, notably solving math, coding, and Ph. D.-level science problems. The chart below shows o1’s performance on competitive math and coding challenges and compares its abilities with human experts on science challenges.

The latest o3 model fares even better, particularly on software engineering and competitive coding challenges.

Other companies have been quick to follow OpenAI, partly because their AI researchers have found that the scaling hypothesis—the view that scaling models will continue to grow their apparent intelligence—is starting to falter. Rumors are rife that OpenAI’s new model, codenamed Orion (think GPT-5), does not deliver the same leap in capability seen in the jump from GPT-3 to GPT-4.

Rather than trying to stuff all the intelligence in a model at training time, AI companies now use chain-of-thought and other reasoning methods at test time (at inference, when you run the model) to boost intelligence and accuracy. I expect this trend to continue. Scaling has always been destined to top out at some point. If only because the economics and practicalities of growing the size of AI data centers are severe limiters. Things are pretty nuts when an AI data center costs $100 billion to build and needs a dedicated, co-located small modular nuclear reactor to feed it the 5GW of power it requires. (This is Microsoft’s plan for their upcoming Stargate data center, which may come online in 2028 or sooner. Why sooner? When xAI announced its Colossus data center, it sent the rest of the industry scrambling to accelerate their AI infrastructure plans).

Types of Agents

Let’s get back to agents. With improving reasoning capabilities, innovators have the building blocks to build impressively intelligent agents. But what kind of agents are we talking about here? Agents will be created for consumers, enterprise workers, researchers, and other applications. For example, ambient agents will work in the background to make life easier for us. Simple agents already exist. HireLogic’s agent for hiring managers and recruiters is an interview assistant that gives detailed feedback on the interviewee (based on whether they demonstrated skills and aptitude in the areas you specify you’re looking for) and creates a transcript and summary of the interview so the interviewer can focus on the interview and not on notetaking. The agent even provides feedback to the interviewer on their technique, letting them know if they forgot to probe the candidate on relevant experience. Anthropic recently announced it added ‘Computer Use’ to its Claude model, allowing its AI to perform tasks using a web browser. These are simple agents, but far more sophisticated, capable agents are coming.

Here are some ideas for agents we might see very soon:

Consumer Agents

• ‍Travel Agent – An interactive AI-generated avatar you can talk to about your travel needs that gets to know you and your preferences over time and helps you research and book trips.‍
• Shopping Assistant – An agent that understands your needs and wants, performs extensive research to find items that meet your expressed criteria for quality, price, form, and function, and once you give the go-ahead, gets the best deal possible.

Enterprise Agents

• ‍Customer Support Agent – An agent that fields customer support calls, books appointments, answers questions, sells spare parts, and performs any needed follow-up, such as sending an email with a PDF of instructions.‍
• Buying Agent – An AI assistant for the purchasing department that helps to negotiate supply contracts, watches the marketplace for price adjustments, and proactively manages inventory levels to meet demand.‍
• Scientific Researcher – An agent that reads scientific papers, posits new theories, designs and conducts experiments (either digitally, in simulation, or physically using robotics), and writes up its conclusions.

Ambient Agents

• ‍Home Manager – An agent that watches over your home. The agent could allow access to your home for deliveries or work performed by certified people. For example, it might guide a plumber to the right room to fix a leaky faucet or watch over the delivery of a piece of furniture, ensuring the furniture is set up in the right room, all packaging is removed, and that the delivery workers are respectful of your home and your property.‍
• Security Manager – An agent that watches over a public space, such as a train station or airport terminal, for suspicious behavior, alerts when somebody leaves behind luggage, and so on.
• Cybersecurity Manager - A swarm of agents that roam a network looking for intruders, monitoring activity logs, and proactively strengthening network security.

Your next coworker might be a machine

Agents will be intelligent but have limited knowledge of the real world. Think of them as intelligent interns that you must guide to help them succeed. These interns will feel, in many ways, just like remote workers. But these digital workers will work for electrons rather than dollars and cents.

Everyone will become a manager of tens or even hundreds of agents. And those agents may, in turn, have sub-agents that work for them. In the same way that we need to learn how to get the most out of the tools we use today and how to encourage the best collaboration and results from human coworkers, we will need to learn how to get the most out of our new machine coworkers. They will need clear instructions and oversight. The companies building these agents must ensure that the rewards for using them significantly outweigh the oversight tax involved.

Agents will collaborate on tasks, too. Imagine you need to create a new advertising campaign for a product launch. You engage your advertising campaign team of AI agents and carefully explain your requirements: Product details, value proposition, market positioning, pricing, target markets, and so on. The agents go away and build a comprehensive set of creative assets, a social media strategy, a broadcast media strategy, a detailed budget, and the projected revenue the campaign will generate. Before the agents return to you, they realize they must get a legal review of the creative assets. So, they connect with agents in the legal department to perform a first-pass legal review. All this occurs in a matter of minutes.

Swarms of agents may work together to achieve a common goal, such as patrolling a network to prevent cyberattacks. We may soon live in a world populated by a trillion intelligent AI agents working in the foreground and background to perform tasks on our behalf. Agents will manage other agents, but that’s still a lot of agents for us to keep track of and provide oversight on. We may see the emergence of a ‘chief of staff’ agent that will become our primary point of contact with our agent posse. This agent would talk to and manage the activities of all the other agents, streamlining the process and making it easier for us. If you’ve seen the movie Iron Man, you’ll be familiar with Tony Stark’s Jarvis AI assistant. Now imagine that Jarvis is your chief of staff AI, overseeing the efforts of many hundreds of agents working beneath it…one that monitors and manages your health, one that manages your calendar, one that does your shopping, and so on. They may all seem like one assistant, but behind the scenes, hundreds of agents might work together on your behalf.

What’s Next?

Tech companies, big and small, are readying their agentic AI offerings. OpenAI will soon launch Operator, and Google DeepMind’s impressive Project Astra research project will likely become a personal assistant service at some point soon. Smaller companies are building agents designed to offload all manner of work from us: Calendar managers, purchasing bots, network managers, and more.

If you want to design and build your own AI workers, tools are emerging to help you do exactly that. MindStudio has launched a solution that makes it easy to build agents using a simple graphical interface. Simple agents can be built without code, while developers can use the platform to design complex behaviors and use their favorite LLM models.

The near future of work will be about human, digital, and robotic employees working closely together to accomplish goals. Each will have unique strengths and weaknesses. Companies that figure out the right balance of employees in their organization, build a culture of trust between their human and machine employees, and use digital employees and robots to amplify and elevate the efforts of their human employees will win in the marketplace.

As Sam Altman has proposed, ultimately, we may see the first unicorn ($1 billion) company with only one human employee, with most of the company’s operations run by swarms of digital agents and robots. A single employee may seem unrealistic, but perhaps a small C-suite supported by thousands of agents and robots isn’t too far-fetched. Time will tell. What is certain is that the workplace of 2030 will look very different than it does today.

And the transformation begins this year. Get ready!