Dear Aventine Readers, 

Since our early days we’ve been following multiple stories about artificial intelligence and its effects, ranging from its seemingly inexhaustible energy demands to how it’s affecting medical research to whether it can achieve consistent and comprehensive human-like intelligence. Today we explore yet another way the technology could transform our lives: the use of AI agents. 

These agents, also known as agentic AI, can be deployed to make autonomous decisions and then act on them without ongoing user input. They are making it possible — both for individuals and companies — to hand over tasks and entire projects and processes to artificial intelligent systems. How revolutionary could these agents be? Their advocates believe that they will be taking on full-time staff roles within a year. But like most AI-systems, agentic AI involves significant risks, and those who know the technology advise caution. Read on to learn more about what these agents can do, and how they can be deployed. 

Also in this issue: 

Thanks for reading! 

Danielle Mattoon 
Executive Director, Aventine

Subscribe

Subscribe to our newsletter and be kept up to date on upcoming Aventine projects

If there’s a defining technological trend so far in 2025 it is, for better or worse, AI agents. 

We previewed the technology in January and since then companies have increasingly begun experimenting with what so-called agentic AI can do. A recent survey of business leaders by KPMG showed that 65 percent of companies are testing the use of agents to help automate tasks such as call center work, recruitment and day-to-day administrative tasks, up from just 37 percent at the end of 2024. OpenAI CEO Sam Altman has gone so far as to write that, “in 2025, we may see the first AI agents ‘join the workforce’ and materially change the output of companies.” Anthropic, the $60 billion artificial intelligence startup, thinks those agents could take on full-time staff roles within the next 12 months.

But how, exactly? 

As it is with almost every nascent technology, it can be hard to know what to make of agentic AI. These agents, which use AI to help autonomously make decisions and take actions, can at once sound transformative, mysterious, a little terrifying, impossible to implement and also … sort of banal. Marketing campaigns and media coverage seem to have, perhaps unhelpfully, leaned on the idea that agents will become our personal assistants, rearranging our schedules, booking our flights and making our restaurant reservations, rather than fundamentally transforming the day-to-day operations of businesses in the way that Altman suggested may be possible.

“People are definitely confused,” said Sanjna Parulekar, senior vice president for AI product marketing at Salesforce. “AI agents are not different [from] any other kind of technology hype cycle that we've ever seen. Everyone starts talking about this new thing, and then everyone wants to pretend that they know what it's about, and then they quickly realize that their understanding is about a mile wide and an inch deep.”

There is plenty to be confused about. What actually are agents? Where should businesses be using them? What are the risks? And are late adopters in danger of forever missing the boat? Aventine spoke with experts at companies that are developing AI agents in order to answer these questions.

The many faces of an AI agent

Most of us have now made use of chat-based AI models like ChatGPT, Google’s Gemini and Anthropic’s Claude. This is an active experience: You prompt the AI with a question and it comes back with a response. In contrast, AI agents use multiple tools — which might include large language models but also systems that allow the agents to access data and interact with other pieces of software — to facilitate tasks without the user being involved at all. “It's automatically doing things on your behalf without you needing to go to it,” said Mike Knoop, a co-founder of the automation software company Zapier.

These agents can vary greatly in complexity. At one end of the spectrum, an agent might be just a string of simple tools and rules chained together to automate a particular workflow. Knoop gives an example of how an AI agent could be deployed to execute an email marketing campaign: A company might want to send emails to newly registered customers, but only those with, say, work email addresses. An AI agent could be built that uses simple rules-based approaches to identify relevant email addresses before turning to a large language model to generate a customized message for each selected customer. Finally, the agent might use an application programming interface (API) to send that message using an email tool like MailChimp. All a worker would need to do to make this happen would be to set up some simple rules, write prompts for the LLM to help it compose the email, and then leave the agent to do its job. “You're plugging AI into that process somewhere in order to supercharge it,” said Knoop.

At the other end of the spectrum, agents can assume more of a decision-making role, explained Ruchir Puri, chief scientist at IBM Research. “They take a complex task, they kind of break that task down, they map it to tools, they execute it, they get the results together, they sort of integrate it, they reflect on it and then, based on that reflection, they'll go and modify their workings,” he said. “They may replan. They may remap [how they use] the tools.” He explained that IBM is developing bots, for instance, that can help monitor, manage and optimize a company’s cloud computer spending, a practice known in the IT industry as FinOps.

Richard Riley, general manager of Power Platform, the Microsoft system that allows its customers — through the use of agents — to build new business tools without writing traditional code, compares the universe of AI agents to apps. “You've got a vast range of types [and] capabilities of apps,” he explained. Over time, he predicts, we’ll come to think of AI agents as tools that you pick and choose to perform certain types of work, just like you pick and choose specific apps to get certain things done. He also pointed out that it is helpful to think beyond the consumer examples of, for instance, having agents organize calendars or book flights, and to focus on more mundane business tasks that could be automated, such as processing invoices or monitoring and routing customer emails. “It's kind of the boring stuff, but it's the stuff that we all have to do day in, day out,” he said. “And honestly, it’s where the money is made.”

How should businesses think about using agents?

Like many other forms of IT infrastructure, there isn’t really a copy and paste approach for companies looking to use agents. Each company needs to ask itself a series of questions to determine how or if it makes sense to use the technology, and to consider many of the same issues that any other AI deployment would require — clearly identifying a problem, understanding available data, making sure that data is usable, and so on.

First, you’ll want to identify places “where you have repetitive actions happening,” suggested Knoop. Once you’ve identified a specific business process, “dissect it, and then figure out where you can start to apply the technology to it,” said Riley. This works best if you involve your subject matter experts “from the very beginning,” said Puri, because they are intimately involved with the work and will understand the data sources available to help an agent perform a task effectively. On the topic of data: Parulekar points out that while customers often know that access to data is an important part of deploying agents, they often don’t realize how poorly structured or insufficient in volume the data inside their company is, which can hold up the deployment of AI of any sort, including agents. “"These agents can be incredibly powerful if they have the right data, but they can't fix your data problem overnight,” she said.

Almost as important: Don’t get carried away. “You're not going to walk in and just completely replace, like, your [entire] recruiting process,” warned Riley. While it might be tempting to tackle a big challenge that has historically been hard to solve, “pick something simpler to prove the technology,” said Puri. “Get some experience with it. Get hands on with it, and really play with it.” Riley likened the idea of tackling mission-critical processes as a first use case for agents to “getting into a Ferrari when you've just passed your driving test.” You can do it, he said, but you would “probably want someone sitting next to you making sure that you press the pedals the right way.” Both Knoop and Puri also argued that now might be an important moment to formally train team members so that they better understand how to use AI inside the business.

What happens next depends on what exactly a company is trying to achieve. Riley explains that some of the agents Microsoft is building will simply become part of its products — say, an agentic meeting facilitator in Teams that can prompt latecomers to join by sending them an instant message and then take notes, share appropriate action points from the meeting after it has ended and so on. Other third-party companies, such as Zapier and UiPath, offer similarly off-the-shelf agents to facilitate processes between different systems — say, Google Workspace or Microsoft 365 with platforms like Eventbite, Hubspot, Slack or other types of software.

In many cases, though, a company might need to build an agent specific to its own requirements. Riley points out that while many business processes are similar between organizations, the on-the-ground reality can be subtly different, often requiring customization. One solution is for in-house software developers to build the product from scratch. Another is to use so-called low-code or no-code approaches — tools that allow people who can’t write code to build agents by defining workflows, making rules, choosing specific tools and granting access to data. IBM, Microsoft, Salesforce and Zapier all provide products that allow for this approach, as do many other companies. 

Ultimately, it may just boil down to giving it a try. “It's not this pie in the sky [idea where] this agent's going to solve everything you've ever wanted,” said Parulekar. “[But if you have a] job to be done, have an agent take a crack at it and … track ROI the same way you would with your team.”

Could something go wrong?

The short answer to that question is: yes. The longer, more nuanced answer is that there are obvious potential issues relating to hallucinations, bias, privacy and security, and while companies building agents are trying to put guardrails in place, a lot will ultimately depend on the risk appetite of the company considering using agents.

Any agent built on a large language model is at some risk of suffering from hallucinations and biases. “One of the downsides of AI is that it can be inaccurate and wrong,” said Knoop. If an agent is interacting directly with a customer and says something problematic, it could be reputationally damaging; if it has permission to perform financial transactions and does so incorrectly, it could lose money for a company. Knoop said that the technology “just isn't reliable enough” to put in front of customers if it’s not supervised by human employees, and generally encouraged the use of what he called a “human curation layer” with all agents. Parulekar, meanwhile, said that if companies do want to try to use agents in customer-facing situations, they should do so initially in very limited ways — perhaps by replacing answers to a very specific type of question in a customer service chatbot with agentic responses while answering the rest of the questions with existing software.

Puri thinks the most successful agents will have safeguards baked in. For example, an agent designed to work in an HR setting might need to check its actions at every stage against a checklist to make sure that it’s always compliant with company rules and legal requirements. Salesforce offers its customers a platform to “battle test” its agents before they’re deployed in order to unearth problematic behavior, while Microsoft’s agents alert their owner if they struggle with a task rather than attempting to solve the problem through guesswork, Riley explained.

As for privacy and security, the potential for problems is significant. At South by Southwest in Austin, Texas, last month, Meredith Whittaker, the president of the Signal Foundation and chief adviser to the policy think tank AI Now, said there was a “profound issue” with granting autonomous software access to sensitive data. Many of these tools may need to move data back and forth across the cloud to perform tasks, and there is currently no way to do so while retaining data encryption.This seems particularly problematic when dealing with an individual's data — especially if that includes calendars, email, messaging and other personal information.

Riley said that it’s possible to build agents with strict limits on what data they can and can’t access, whether they have permission to read or write inside documents and so on, much like the data restrictions that can be added to a person’s user account on a website. But there is still an undeniable risk associated with software designed to work autonomously — an agent could be hacked to act maliciously, say, or it could leak sensitive data. One important consideration here is how agents are deployed. Knoop notes that while they are often rolled out from the top in organizations because of an efficiency mandate in the C-suite, sometimes they are used by individuals bottom-up to solve specific problems, a phenomenon referred to as “shadow IT.” In the latter case, it’s far more likely that data governance practices aren’t properly adhered to.

Concern about these issues seems to be “on a spectrum,” according to Parulekar. Knoop said that at this point, the majority of mid- to large-size companies have taken a view on AI and data governance, with preferred approaches and vendors based around their own risk tolerance and the degree of regulation in their industry. In smaller companies, he said, there is “really not as much concern on the security and the privacy side of things,” and instead there is more emphasis on what agents can do, and how reliably they can do it.

What if my company isn’t ready to use agents?

With any ascendant technology — or at least, the appearance being ascendant — there is always a nagging concern for institutions that haven’t yet adopted: Are we being left behind?

The answer is no, at least not yet, according to several experts who all agreed on this point. “We are not in a world today — with the current maturity of AI, the reliability of AI — to say that your business is going to fail if you don't adopt [agents] today,” said Knoop. “The technology is not good enough.”

But one thing the experts also agreed on was that thinking about using this sort of AI, coming to grips with how it works and understanding its capabilities, is an important part of preparing for the future. “You need to develop that muscle,” said Puri. “Developing that muscle takes time, just like, you know, exercise takes time.” 

“You're not gonna get out-competed by a company that's all in on AI right now,” said Knoop. “Could I see that happening in the next five to 10 years? Absolutely.”

The dire wolf went extinct about 13,000 years ago. These two were produced in a lab. Colossal Biosciences

Scientists have brought extinct wolves back to life, sort of. The dire wolf — a large canine that was contemporaneous with saber-toothed cats and other ice-age creatures — went extinct about 13,000 years ago. Now a company called Colossal Biosciences, which made headlines in 2021 for announcing plans to bring the woolly mammoth back from extinction, claims to have created three dire wolves, which are — according to the company — now roaming around a secret 2,000-acre reserve somewhere in the northern U.S. But are the animals true dire wolves? It’s an important question, because the approach the company took, Wired reports, wasn’t like the cloning from DNA used in Jurassic Park, which, in both the novel and the film, created exact replicas of dinosaurs that roamed the Earth more than 140 million years ago. Instead, Colossal sequenced the DNA of dire wolf fossils, identifying key features in the genome relating to attributes such as size, build, fur color and so on. Then the company edited the DNA of a regular gray wolf to include those features, making 20 edits in total — the most performed on the DNA of any one creature, according to the company. The result looks more akin to the dire wolf than the wolf it’s based on: broader shoulders, a wider head, and its fur is white rather than gray. But just because it looks and behaves like a dire wolf, does that make it a dire wolf? The company claims that it does, but concedes that taxonomists would likely take a different perspective, because speciation is defined by a combination of both physical characteristics and genetics — and the new dire wolf is by no means a full genetic replica of the extinct version. Still, the company sees the advance as an important first step toward rescuing a species from extinction, and partly chose the dire wolf because it believes it can transfer the technology to efforts to conserve related species, including the critically endangered red wolf. 

Could AI reduce CO₂ emissions more than it increases them? In a new report about artificial intelligence, the International Energy Agency (IEA) brings new detail to an increasingly familiar story: Electricity consumption by data centers is expected to more than double by 2030, to 945 terawatt-hours per year — about the same as the annual consumption of Japan — largely driven by increasing demand for AI. Yet the same report also highlights how AI itself could help reduce emissions by helping solve all sorts of energy and climate problems. AI could make power plants more efficient, spot methane leaks, reduce energy consumption inside buildings, optimize energy use across the grid, improve efficiency of logistics networks, identify more climate-friendly industrial processes and so on. (The industrial process example is particularly interesting, and The Economist has taken a close look at how AI can be applied to heavy industry, a part of the economy that is notoriously difficult to decarbonize. The IEA predicts that widespread adoption of AI in this sector alone could save around 2,200 terawatt-hours of energy per year by 2035 — about the same as the current annual energy demand of Mexico.) Collectively, according to the IEA, projects using AI could reduce global emissions by as much as 1.4 billion tons per year by 2035, which would be up to three times as much as the annual emissions from data centers by that date. Still, as MIT Technology Review warns, this report speaks of the promise that AI offers in cutting emissions, which doesn’t solve for the fact that AI is contributing to carbon emissions right now. Without market or regulatory mechanisms in place to ensure that AI is implemented responsibly to cut emissions, the potential outlined by the IEA is merely that. 

A new understanding of pain could help unearth better analgesia. A team of researchers from Stanford University in California has used synthetic biology to recreate in a lab the sensory pathway that transmits feelings of pain to the human brain, an achievement that will help the development and testing of new kinds of pain relief. The team, whose work is published in Nature, transformed human skin cells into stem cells, the cells in the body capable of turning into virtually any tissue. They then used chemicals to coax those stem cells into four different types of cells representing different parts of the nervous system, from the sensory neurons in the skin through those in the spinal cord, and two found in the brain, including the cerebral cortex. These were lined up next to each other and grown over 100 days into a synthetic representation of a human brain circuit just under half an inch long called an “assembloid.” The researchers were then able to observe how these neurons fired and passed on waves of electrical activity in response to stimuli along the circuit — for instance, firing when the sensory neurons were exposed to capsaicin, the molecule that gives chilli peppers their heat. The researchers told the Financial Times that the approach will expedite the development and testing of new forms of pain relief without the need for human or animal testing, potentially leading to new drugs that dampen neural activity associated with pain without the addictive effects of many forms of pain relief. 

Failure to communicate, from Science
3,700 words or about 15 minutes

There’s a recurring problem plaguing geoengineering efforts to modify our skies and oceans in attempts to fend off climate change: Scientists don’t explain to the public what they’re doing, leading to a public outcry that can shut down the experiments.This happened with a plan to test stratospheric aerosol injection in the skies above Sweden, with a proposed trial of marine cloud brightening off the coast of California and with an ocean alkalinity enhancement project in the sea around Cornwall, England. Each trial was forced to pause due to public outcry. But, as this story from Science explains, the public doesn’t necessarily oppose such projects by default. In fact the first team to successfully test a solar geoengineering approach outdoors — a cloud brightening trial off the northeastern coast of Australia — organizes its research with the public in mind, claiming that the approach helps identify experimental risks that might not otherwise be considered. As we look toward a future in which geoengineering is likely to be a component in the battle against climate change, researchers may be able to learn a lot from that Australian project about how to move geoengineering projects forward. 

How Bryan Johnson, Who Wants to Live Forever, Sought Control via Confidentiality Agreements, from The New York Times, and When I’m 125? from Coda
3,200 and 3,900 words, or about 27 minutes in total

Longevity has long been a preoccupation of the Silicon Valley elite, and if there is a poster child for the idea that aging can be defied it is Bryan Johnson. He lives according to a regimen few people could replicate that includes eating three meals between 6:45 a.m. and 11 a.m. consisting mostly of vegetables, nuts, berries and seeds. He takes over 100 pills a day, has received transfusions of his son’s blood with the goal of pepping up his own and has established a self-proclaimed religion based around his mantra “Don’t Die.” He also runs a startup, Blueprint, that promises to share the secrets (and supplements) of achieving a longer life. But as these two stories explain, all may not be well inside Johnson’s empire. The New York Times reports that Johnson makes frequent use of nondisclosure agreements to keep some of his more unorthodox practices under wraps. Meanwhile, his longtime longevity doctor left Blueprint last year over concerns about the company's supplements. In Coda, J. Paul Neeley writes a first-person account of being part of a so-called self-experimental Blueprint trial, which he and fellow study participants each paid $2,000 to join. Many of those involved in the trial experienced negative side effects that included prediabetes, reduced testosterone and tinnitus— reactions that the company discouraged participants from revealing. Ultimately, Neeley — himself a longtime life optimizer — wrote that the experience made him think more deeply about the longevity movement. "I’ve ... started to think that calls to live forever are perhaps misplaced," he concludes, "and that in fact we have evolved to die."

The Worm That No Computer Scientist Can Crack, from Wired
2,400 words, or about 10 minutes

The C. elegans nematode is a lesson in biological efficiency. The worm is made up of fewer than 1,000 cells — about 300 of those being neurons — and yet it can find food, eat, move and reproduce. The creature’s simplicity has made it the focus of an open source movement dedicated to replicating the worm in software down to the individual molecule, which would make it possible to create a model that could help us understand the most basic biological functions in exquisite detail. Perhaps it could even help us understand how our brain works, this story suggests, by exploring at a very basic level how nervous systems work from the ground up. Yet it turns out that creating a simulation of even the simplest creature is incredibly difficult. Just amassing sufficient data through genetic imaging technology to accurately reverse engineer this little worm is predicted to take 10 years, cost tens of millions of dollars and require as many as 200,000 sacrificial worms. Whether that’s a sensible use of time and resources is very much up for debate. The striking thing about this story is that it reveals the complexity of even the simplest form of life, and how attempts to understand the world through computational simulation, while important, can also be fiendishly difficult.