What do we mean when we say AI agent?
by Gil Dibner Partner at Angular Ventures. | Originally published on The Angle newsletter.
Guest post by Gil Dibner Partner at Angular Ventures. | Originally published on The Angle newsletter, make sure to go sign up ❤️
In our conversations with startups these days, the word “agent” has rapidly become one of the most commonly used and most confusing. There does appear to be a commonly accepted technical definition of “agent” as applied to software (or “agentic software”). This definition refers to any system that is both autonomous and goal-seeking. Those two attributes appear to be essential for something to qualify as an agent. Beyond that, however, there is a lot of confusion in how the term is actually used.
A quick search for agent taxonomies reveals various taxonomies that can be used to categorize agents by various attributes. Agents are sometimes categorized by their logical approach to decision-making (simple-reflex, model-based, goal-based, utility-based, or learning), by their degree of autonomy (fully autonomous, semi-autonomous, manual), by their internal complexity (single-agent, multi-agent), by their architecture (deliberative, reactive, hybrid), or other by any of several other dimensions that can get increasingly specific.
But in the context of most of our discussions with entrepreneurs, these definitions and categories are rarely mentioned. In our experience, when a technologist today says “agent” he or she is usually using the word in one of four ways:
Anthropomorphic (or literal) agents are AI agents that have a visible digital presence that users can directly interact with such as virtual assistants. Crucially, their interface is designed to mirror that of a human “agent,” usually via text or voice chat. These agents perform tasks visibly and predictably, allowing users to observe and control their actions in real-time. Good examples of visible agents would be customer service chatbots, robotic SDRs, virtual market research panels, or virtual travel agents.
Functional agents are specialized AI agents that perform specific tasks behind the scenes, often without direct user interaction. They focus on one or a few functions, optimizing workflows and enabling efficiency in automated processes. They can replicate human labor or other software systems, but they are typically autonomous enough that they are not interacted with directly. Examples of functional agents would include sales lead enrichment, marketing content generation, or procurement management.
Adaptive agents are highly versatile AI systems designed with broad capabilities to handle a wide variety of tasks, including those not known or defined in advance. These agents autonomously adjust their strategies, leveraging advanced learning algorithms and adaptability to tackle new challenges as they arise, making them well-suited for dynamic and unpredictable environments. The best examples of adaptive agents are usually offshoots of the foundational LLMs produced by the hyperscalers such as OpenAI, Google, or Meta. These foundational models (and companies building on them) can engage in a wide-range of tasks without any a priori configuration.
Composable agents are autonomous, modular building blocks within complex software systems, distinct from traditional API microservices by virtue of their ability to make decisions and manage tasks independently. Designed to be invoked and combined by other agents or function calls, composable agents provide reusable functionalities that can adapt to various contexts and support larger workflows without continuous oversight, allowing systems to scale and flexibly orchestrate multiple agents. For example, in a cloud-based architecture, a composable agent might autonomously monitor and handle data encryption, deciding on encryption levels based on data sensitivity, rather than merely responding to API calls. Similarly, in an e-commerce platform, a composable agent for order tracking could proactively manage tracking updates and optimize delivery routes in response to varying conditions, enabling other components to leverage these functions seamlessly.
We see tremendous potential across all four types of agentic systems - both on the application and the infrastructure layer. We’ve already invested in one functional agent company (still in stealth) and two infrastructure companies that can be used to assemble composable agents into next-generation agentic architectures.
But given how quickly this space is evolving, we are far from certain that the categories above are correct or even useful. And so I’ll throw the question back to you: do these categories make sense? Are we missing something? Is there a different definitional framework we should be applying? Most crucially - if you are building either an agentic application system or infrastructure for enabling such systems, please reach out. We’d love to learn about how you see the world.
