[Editor’s Note: Agents of Transformation is an independent GeekWire series, underwritten by Accenture, exploring the people, companies, and ideas behind AI agents. Join us Tuesday, March 24, for our Agents of Transformation event in Seattle.]
Last year, AI-powered sales platform Nooks didn’t use multiple AI agents, instead relying on pre-trained alerts and models to help customers fine-tune their sales strategies. But that has changed a lot.
“We’ve injected ourselves into almost every part of the stack since then,” said Nooks co-founder and CTO Nikhil Cheerla, speaking at an event the company hosted in Seattle in February.
The rapid adoption of Nooks reflects the growing attention to static AI agents – tools designed to perform a single task well by integrating models with domain-specific, workflow-, and context-specific data.
General purpose AI models can quickly generate text, write code, and summarize reports. But they don’t have the ability to solve industry-specific tasks. And that’s where direct agents come in.
For this installment in our In the Agents of Transformation series, GeekWire explored the growing trend of virtual AI agents, and the huge potential for startups.
“Major AI platforms may become broad distribution engines of intelligence,” Madrona investors Sabrina Wu and Vivek Ramaswami wrote in a recent analysis of the AI landscape. “But specialized companies will continue to emerge by finding difficult parts in certain domains.”
Jerry Zhou, the CEO of Seattle tech startups, described direct AI as “the shift from tools to agents.” Supio software helps attorneys quickly organize, search, and organize case-related data.
“It’s not enough for AI to generate insights — it needs to work within real workflows and take action,” Zhou said. “In legal terms, that means turning complex data like medical records into verifiable, systematic outcomes advocates can rely on without guesswork.”
New technology helps startups uncover customer value. Prophetic, a Portland, Ore.-based data intelligence platform, has trained its AI on more than 20,000 metro area codes in the US. “That’s the true power of direct AI.”
The shift is drawing attention from investors like Mia Lewin, a Seattle-based veterinarian who recently raised a $5 million seed fund for TheFounderVC, her new company focused on vertical AI startups.
“We expect this space to include more than 300 unicorns over the next decade, with the first Vertical AI IPOs coming to market within three years,” Lewin said.
Speaking last month at Nooks event, Pulumi CEO Joe Duffy described how Pulumi’s AI agent, Neo, helps companies automate cloud infrastructure tasks such as optimizing costs and ensuring compliance. The goal of Neo, which was founded last year, was to create an AI agent that can do everything a human infrastructure engineer can do – not just to answer questions, but to take action in all complex systems.
“One of the special features of a direct agent is that you can go deep into one domain,” Duffy said. “And that domain is not just LLM tokens. It’s much more complex than that.”
Building these systems requires more than a model. It requires what some call an “agent harness” — the surrounding infrastructure that helps organize tasks, get context, and validate outputs, Wu and Ramaswami noted in their post.
Vertical AI agents are already performing different types of manual work for themselves – they go beyond traditional software-as-a-service tools.
“Turning the context of a workflow into a killer opportunity for direct AI agents, and what will separate the winners from those who only produce content or recommendations,” said Doug Tallmadge, CEO of Seattle marketing AI startup Gradial.
Startups pairing static AI agents with robust contextual data may pose a threat to incumbents. Cheerla, CTO at Nooks, said a company like Salesforce has billions of data points — “but they don’t know what’s good and what’s bad in that data.”
“The way we’re trying to design Nooks is to collect the highest quality data, to get the full context that led to the decision,” Cheerla said.
Nooks agents manage the end-to-end sales workflow, including identifying accounts, finding contacts, writing emails, and assisting live call responders. They can be manually invoked, deployed in bulk, or run in the background, and are designed to work in collaboration with human users.
The next class of virtual AI agents can go beyond simple task execution. One emerging trend is agent-to-agent collaboration, where multiple systems work together to solve complex problems.
“You can think of a bunch of agents working together to do something,” Duffy says, drawing parallels to his earlier work designing distributed systems.
Another change is to active agents – systems that not only respond to instructions, but initiate actions themselves. However, that change may take time. Even as agents become more powerful, companies are moving cautiously when it comes to handing over control.
Duffy talked about the “slide to autonomy,” a term coined by AI researcher Andrej Karpathy, which ranges from fully human-controlled systems to autonomous agents.
For low-risk tasks – such as cleaning up unused cloud resources – companies can allow agents to work independently. But for high-level actions – such as deploying production infrastructure – human observation remains important.
“You have to start building trust and building quality in the programs you build,” said Duffy.
Powerful direct agents have begun to reshape how companies organize their teams. Cheerla described a shift in the traditional model of engineering organizations, where product managers help share knowledge between engineers and customers. He said that process can be automated by agents, and engineers should connect directly to customers and gain ownership over results.
“You have to remove these pipes and obstacles,” he said.
At Pulumi, Duffy described a shift where every developer effectively leads their team of agents. “Engineers who can think like a product manager and a staff-level engineer can be 100x developers,” he said.
Investors with Bessemer Venture Partners say vertical AI “represents a much bigger opportunity than vertical SaaS ever did,” in part because of how it affects the workforce.
“Unlike vertical SaaS, which often takes up a small portion of Fortune 500 IT spend, Vertical AI impacts directly on the workforce’s P&L bottom line,” they wrote in a blog post.
Sharbani Roy, a vice president at chip design firm Arm who once helped build Alexa at Amazon, offered a different framework for how workers interact with agents: the learner model.
Instead of thinking of agents as automated tools, he encourages his team to ask a different question. “How do you use an agent to help you behave as a student to make yourself better?” he said in the interview. “What did you do this week that you were able to achieve – but better – because you had an agent helping you? How do you make high and high judgment calls?”
