Agents Agents Agents

It's the new buzzword for 2025

It’s the buzzword of 2025.

No longer is simply getting into AI enough, CEOs all over the earnings will start appending the word “agents” as many times as possible in their transcripts. Startups pitching AI will be obsolete, they will now need to pitch “agents” to get funding. We’ve gone into AI fatigue, and need to evolve into the next phase…

So what are agents? Well, last month we talked about examples of generally available agents to the marketplace, such as Windsurf. AI agents are already demonstrating a few attributes that are pretty incredible. For example with Windsurf, one was the ability to plan out an implementation given a prompt, and then the ability to utilize many different tools and take advantage of abilities such as creating files and folders, executing terminal commands, and even self-learning to debug when things didn’t work. There is a lot going under the hood here which we will get to in a bit.

So what can we expect in 2025? Well first the barrier we will have to get past is trust. Why? Because agents can really be effective when they have access to more tools and data, which is where society is going to need a leap of faith.

For example, let’s take the most common use case we hear on podcasts and news articles as what an agent can do…book a vacation

Let’s walk through what it would take to have a successful travel AI agent accomplish this today:

Calendar access, not only my calendar, but it needs the access to the types of travel, events, and hotels to match. The good news? There are already APIs that can access all of these, but you have to give the AI agent yours as well.

Context. How will the AI agent know about who this trip is for, is it for my family or just myself? Is it for business or is it for soul seeking? This is also easy right, much of the context can be included in the prompt, or is it?

Preferences. How will an AI agent know what I like? What kind of activities, food, the type of destination I would like over another one? This one is tricky, as it may require you to give way more data to the AI agent than you’d like. If you connected all your historical trips and itineraries, maybe that would work. Or if you connected your social media accounts? That’s asking for a lot.

Payment options. That’s right, in order to complete the circle of tasks, you need to entrust your AI agent with your credit card! How else would it be able to book all these things for you without putting something down to reserve things? However, how would it know your spending limits, what if it booked first class? Maybe someday…but not yet.

So notice above, in the context of booking a trip with an agent, everything is theoretically possible today. The main issue is I had to give access to my calendar, preferences (which could range from a large set of data sources), and payment options. What if the agent booked something that was super expensive? How would I resolve mistakes?

Now let’s jump back to enterprise, every company right now will be creating agents for all sorts of tasks, and they will demand very similar amounts of access to data, tools, and preferences. As long as we are okay with that, agents will could be successful. Once we cross the line where the level of trust is satisfactory, then these barriers will more and more likely get unlocked.

Let’s talk through a few examples. There will be some powerful use cases unlocked, we mentioned coding earlier, but support agents are now becoming more and more widespread. Consider what kind of access is needed here, there will be historical tickets of course, but the access to immediately reply and solve problems on behalf of customers could also pose risks to companies. At what point will we trust these agents to issue refunds or make code changes? The benefits outweigh the risks here, as companies are putting more and more tickets to AI agents already today, and the cost savings are one axis but fixing customer problems quickly and increasing satisfaction is another. It’s simply worth investing here.

But as we look deeper into other verticals, consider the use case for Sales. Having an AI agent call and email leads for you, and representing the company’s values and messaging, could be great, however what if it slips up and lies about something? What if the agent interferes with an existing deal in progress? Could it collect data here and develop marketing materials as well? As we become more dependent on agents and weigh the tradeoffs of similar human mistakes, we could also end up becoming less creative and less able to adapt to the changing market as well. Could agents end up hurting in some use cases? Again, the benefits outweigh the risks but the key here is we will be iterating and slowly crossing these lines of trust as the technology improves or as startups focus solely on these verticals.

Ultimately, as 2025 progresses, the key progression won’t be whether vertical agents are happening or not, it will be how gradually we are trusting these applications with more and more of our data and how much they can actually access your systems. It may be so gradual that we continuously get used to it.

Successful AI agents will need to have access to all your files and folders, all your data and preferences, and even other accounts and systems that you have access to as well. Are we prepared to hand it over? The answer is yes. It will just happen slowly. Are Enterprise, Federal, and other entities ready to hand it over also? The answer is yes. It will happen even slower. So is 2025 actually the time to use AI agents as a buzzword? I think the answer is still yes, we’ll see it happen to individuals quickly, then slowly with larger organizations. There will be scenarios where agents make mistakes and cause bad press, so buckle up.

The term AI is already obsolete, AI agents is the new buzzword, but where will you draw the line?