Agentic AI is moving fast, but are we ready for it?

"We're all going to be CEOs of a small army of AI agents," predicted Erik Brynjolfsson, director of the digital economy lab at the Stanford Institute for Human-Centered Artificial Intelligence and founder of Workhelix, recently quoted in The New York Times. "We have to think, OK: What is it we really want to accomplish? What are the goals here? And we have to think a little bit more deeply about that than we have in the past."

Also: 15 new jobs AI could create - could one be your next gig?

Brynjolfsson is the ultimate techno-optimist, so his caution toward marshalling armies of AI agents should be heeded. This view was also shared recently by Dario Amodei, co-founder and CEO of Anthropic, who has openly sounded the alarm about the risks of AI. He, too, sees the advent of human-led agentic workforces, particularly within technology shops. 

The first adoption and success stories for AI agents are starting with coding tasks -- but end up rapidly expanding to other roles within businesses as well, Amodei said, speaking with Databricks CEO Ali Ghodsi at the recent Databricks conference. Currently, the emphasis for AI agents means multiple agents working together, relying on extended networks of agents that can communicate with and activate each other to meet the needs of the moment. 

However, they will not be operating autonomously -- even AI's most visible and vocal proponents say there will always need to be humans to steer AI actions. Within an agentic AI architecture, Amodei predicted, "you dispatch a number of agents to do things for you, where you act as essentially the managers for the agents." These could be called "agent fleets," which will evolve to "agent swarms."  

Anthropic's foundational model, Claude, is employed widely across organizations for a variety of purposes, from coding to business functions. 

Also: Anthropic mapped Claude's morality. Here's what the chatbot values (and doesn't)

"Coding is moving the fastest -- it's where we're seeing things first," said Amodei. "It's a foreshadow of what's going to happen across all applications. We started with coding, and at first, people just write code, then there's the autocomplete era, now we're moving into what's being called vibe coding -- where you kind of ask the model to do something, and it's very interactive."     

When it comes to agents, there will always be a need to have a human engineer sitting on top of a hierarchy -- "like they're managing an organization or a company," Amodei said. Human managers still need to intervene and set direction. At this stage, human managers are needed to make sure agents "don't vent private information, and they don't take actions that would give your security team a heart attack," he continued. 

As an example of the influence of AI agents, Amodei pointed to an open-source autonomous AI agent implementation that originated within Block, the company formerly known as Square. The company's AI coding agent, Codename Goose, is seeing applications starting, but now extends well beyond the developer space.  

Also: For AI agents, bite-size 'small language models' could make more sense

"This is the pattern we've seen," Amodei explained. "At first, people directly used the model for coding, and it received a lot of adoption there, something like 75% of Block's engineers. But it gradually became a general tool for coordinating lots of different types of tool use, lots of different data sources. Models that write code can write code for a user, but they can also write code in order to use tools and take actions."

At least 40% of Block's employees -- beyond software engineers -- are now using Goose, Amodei continued.

Anthropic eats its own dog food as well, and has been discovering new applications for its agents -- including creativity research and financial analysis above and beyond writing code. "We have agents that act autonomously, are kind of able to improve the business, help with a new product release, and develop a new marketing strategy," he related. "We've noticed substantial improvement in our own productivity due to these use cases for Claude. If you yourself are not deriving benefit from something, you probably shouldn't be selling it to others."  

Also: AI agent deployments will grow 327% during the next two years. Here's what to do now

Ultimately, agents will evolve to the point at which they could be considered full-fledged "virtual co-workers," though we're not there yet. "This is something that we're working on and gradually introducing the pieces of," said Amodei. 

It will take time before agentic assistants are fully capable of grasping and iterating through the tasks at hand. 

"It's hard to get the balance to train the models to do what they're being told to do, and the models being overeager and doing more than you expect them to do," he cautioned. "It requires a fine sculpting of the models, following instructions, following human intent, across coding, tool use, various other applications like customer service. That's something that takes a lot of iteration to get it right, and we're still not perfect, but I think we've improved the process quite substantially." 

Want more stories about AI?Sign up for Innovation, our weekly newsletter.

Artificial Intelligence

The best AI for coding in 2025 (including two new top picks - and what not to use)

I tested 10 AI content detectors - and these 5 correctly identified AI text every time

The best AI image generators are getting scary good at things they used to be terrible at

Looking for an AI-powered website builder? Here's your best option in 2025

Use AI at work? You might be ruining your reputation, a new study finds

• The best AI for coding in 2025 (including two new top picks - and what not to use)
• I tested 10 AI content detectors - and these 5 correctly identified AI text every time
• The best AI image generators are getting scary good at things they used to be terrible at
• Looking for an AI-powered website builder? Here's your best option in 2025
• Use AI at work? You might be ruining your reputation, a new study finds