The future of AI agents, according to the builders who will take us there
Vertex Ventures US
We partner with and invest in early-stage founders building infrastructure and SaaS companies.
Welcome back to Vertex Angles, the weekly newsletter from Vertex Ventures US . We’re a boutique venture capital firm, investing in exciting companies across software infrastructure, developer tools, data, security, and vertical SaaS. If you prefer e-mail, subscribe here to get Vertex Angles in your inbox every week.
In this week’s Pi Day edition of Vertex Angles, investor Sandeep Bhadra turns back the clock to a recent dinner with builders in the AI agent space — and, with the benefit of their insights and operational expertise, gazes into our agentic future.
This may be the year of the AI Agent, but the term “AI agent” itself has quickly and easily proven to be the most overused and overloaded term in the technology industry today. The term has found itself broadly applied to a whole variety of different types of AI-powered software, with little nuance.
Recently, Vertex Ventures US gathered up AI agent engineers and product managers from frontier labs and the broader AI ecosystem for a dinner to discuss the state of AI agents today, the businesses they’re most likely to disrupt tomorrow, and the technical challenges to wide adoption along the way. I found it to be an enthralling, deeply technical discussion that shed a lot of light on the nascent market for AI agents, and led to some surprising revelations about the technology and its limitations.
First, we came to an agreement on definitions before diving deeper. We agreed that the common building blocks of AI agents are self-reflecting LLM models, instructions, memory and runtime harnesses/tools to interface with the real world.
Attendees also concluded that AI agents today exist on a kind of spectrum, with creativity on one side and autonomy on the other. Under the most expansive definition, an AI agent should be able to learn, plan, decide and act, ?exhibiting the higher level functions of a gifted knowledge worker. It should exhibit a specialized intelligence in its specific domain, showing up as a sliver of the incoming singularity of AGI. Towards this lofty goal, the frontier research labs are focused on demonstrating ever-higher levels of creativity and sophistication in domains like scientific R&D, mathematical problem solving, or software and data engineering.
At the other end of the spectrum, agents in commercial use are focused on executing well-defined but complex workflows with autonomy and accuracy, while building trust and rapport with their human co-workers.
Looking at the Vertex portfolio, products like Evisort (acquired by Workday last year), Tulip Interfaces , Metaview , Cyberhaven , and SPRX are already showing what’s possible with simple agents automating workflows in fields including recruitment, manufacturing , data security, and tax preparation. Those kinds of products are driving capital-efficient growth, helping each of their human operators do more in less time.
In the course of our conversation, we made the (somewhat unsurprising) observation that researchers value sophisticated models with reasoning and creativity, but product teams are happy with simple AI agents that act as diligent independent workers that can be trusted to execute perfectly and collaborate well with humans.
More importantly, over the past several months, we’ve seen an increasing pace of startups going from $0 to $3 million in run rate (annual recurring revenue is not quite the right term here, as many of these agentic app company revenues are services revenue replacements) in less than 6 months. These companies are AI-native, applying AI-driven efficiencies in every corporate function. Often, even at the seed stage, these are solo founders or two-person teams who have a fully working prototype and early customers. The golden age of AI services-as-software is here, decimating Baumol’s cost disease in stagnant industries and taking a bite out of the trillion-dollar professional services industry.
The future of AI agents and infrastructure
We see this tide of agentic applications as a force pulling forward infrastructure spend and innovation too, as the industry rallies to build the common elements that underlie the whole movement. ?
Today, thanks to workflow orchestration software like Orkes (a Vertex portfolio company), Conductor or Temporal, developers are already stringing together agentic apps with existing microservices with LLM modules. AI-first workflows software like LlamaIndex and Langchain streamline the data ingestion and retrieval process by providing comprehensive API calls for every step in the RAG pattern or by connecting to external tools.
If AI agents are the brains, they need a runtime harness to embody them in practical applications, presenting another area for infrastructure innovation. Github Copilot, Cursor or Lovable – which are automating coding for expert developers and novices – are building intuitive user interfaces combined with the runtime harness of an IDE (integrated development environment).
Looking once again at the Vertex portfolio, this trend puts wind in the sails of Gitpod and Docker, Inc . Similarly, where Hasura 's core API-access-for-databases product continues to grow strongly, its seen strong uptake of its new PromptQL LLM-access-for-databases products.
Finally, in a world where more code is written by bots than by humans, this tide of agentic apps creates new challenges for IT teams to test, secure and maintain these apps, our dinner guests agreed. That sentiment, in particular, was echoed recently by engineering leaders at infra.nyc , a meetup for infrastructure software builders hosted by my Vertex colleague Megan Reynolds .
What does the AI agent ecosystem still need?
At the dinner, we briefly discussed techniques to instruct these agents as well. As agents have gotten more autonomous and easier to reason with, instructions have evolved from simple commands to elaborate discourse around planning out projects. Discussions with the AI agent are the best way to effectively communicate requirements, prepare a reviewable workplan, present intermediate checkpoints/results, check for correctness, receive nudges/feedback, and iterate towards a desired result. If all of these sound like product management/managerial tasks, perhaps the near state of knowledge work is to become an excellent manager of models!
Surprisingly for this group of smart engineers, there was unanimous consensus that the biggest roadblock today to effective practical agentic design is the lack of adequate grounding and eval datasets.
As newer “inference time” agents like o1, o3, Grok3, Deepseek, etc work on multi-step reasoning tasks, it becomes important to actually evaluate the quality of intermediate reasoning steps, not just the final answer and prevent reward hacking or cheating. Combining reinforcement learning with careful oversight through evals and data curation helps build agents that not only speak fluently, but also think more effectively. We believe that observability and security of agents (grounded in these evals and profiling) are a great opportunity for infrastructure investment as well.
Many of the folks present were involved with the MLOps community and the discussion naturally flowed to how agent building can be made more approachable to developers. Do agent frameworks make sense and if so, what abstractions are useful? To wit, what does the React for agents look like?
For many sophisticated users present, the answer was just code, especially in the age of vibe-coding with Windsurf or Cursor. Do frameworks even make sense when your AI copilot/coding agent is writing huge chunks of software in dialogue with a human developer? In other words, is it not the case that coding agents make all other kinds of agentic software design easy for all developers?
These questions are difficult to answer now, but we believe that some of the smartest developers working today are going to build the solutions and help usher in the next generation of AI agents.
The latest news from the VVUS network:
Vertex portfolio job of the week: PromptQL Staff Engineer at Hasura
Hasura is growing its core PromptQL team, working on redefining how AI systems should access data in a business context with an emphasis on visibility, steer-ability, and repeatability. While experience working with LLMs is a big plus, Hasura values strong foundational skills and the ability to learn quickly.
For more startup jobs from across the Vertex Ventures US portfolio, check out our jobs portal.