Transcription
Welcome back to the AI daily brief. Today we are talking about one of the big themes, maybe the big theme in AI right now, which is, of course, agent transformation, agent adoption, the agentification of everything.
Now, since almost the moment that ChatGPT came out, people have been excited about the possibility of this agentic era. Agents represent the idea of an AI that has greater capacity and can do more than just make you more productive, but in fact, fundamentally change the equation on the amount and type of work you can do, or that your organization can accomplish as a whole. Still, what exactly we mean when we say agents can still be a little abstract and difficult.
And so, for this piece, we're building off of the inspiration of a recent piece that appeared in The Information called “The Seven Kinds of AI Agents.” We're going to look at both that as well as a couple of other frameworks to help you wrap your head around the different ways that you might think about categories of agents as a way to inform your personal or organizational agent strategy.
Now, one quick note before we dive into these definitions: Agents are officially and very much in the realm of the now. I shared these statistics before, but in the most recent KPMG pulse survey, they found a massive increase in full enterprise agent deployments. The percentage of organizations that had at least some agents fully deployed, i.e., out of and past the pilot stage, tripled from 11% to 33% between Q1 and Q2. That followed a jump in pilots from 37% to 65% between Q4 and Q1. In net, 90% of the organizations that KPMG surveyed said they were past AI agent experimentation and actively into either pilots or deployments, meaning that this is here and happening right now.
So before we get into a few different ways to break down subcategories of agents, let's talk about the broadest possible definition. If you've listened to me frequently, you'll know that I actually don't care all that much about hyperspecific definitions of agents. I think a lot of the discourse around whether something is an agent or an agentic workflow or an automated workflow kind of doesn't matter. And the reason that it doesn't matter is that I think, in this circumstance, the common knowledge and intuition about the difference between an agent and other types of AI is actually more accurate, or at least more functionally useful, than those highly specific and technical definitions. Basically, if you talk to your average business person or person who works for a big company or corporation about what an assistant or a co-pilot style AI is versus what an agent is, they'd probably give you some definition that roughly said, "assistants are AI that I use to do things," whereas "agents are AIs that do things for me." And I think, broadly and directionally, that is the right dividing line.
So what then is the use of this sort of deeper dive that we're about to go through? Well, even if that is the broad category dividing line between assistants and agents, there still are many different types of agents. And understanding different ways to think about them can be really useful as you're figuring out your agent strategy.
Now, broadly speaking, there are a couple of different ways that people will try to organize agents. The first is based on their functionality. The second is based on their focus.
So, in terms of the definitions that are based on functionality or how they operate, you'll often find this list of six or seven different agents that include simple reflex agents, model-based reflex agents, goal-based agents, utility-based agents, learning agents, hierarchical agents, and sometimes the last that they'll add is multi-agent systems. Here's how AWS defines these things: A simple reflex agent is basically exactly what it sounds like. It operates based on predefined rules and the immediate data it has access to. As Amazon puts it, it will not respond to situations beyond a given event-condition-action rule. These agents, then, are well-suited for very simple tasks that don't require a ton of training. The example that AWS gives is an agent that resets passwords by detecting specific keywords in a user's conversation. Some other use cases for this type of agent from DigitalOcean include automated sprinkler systems that activate based on smoke detection, email autoresponders that send predefined messages based on specific keywords or sender addresses, and things like that.
Next up, we have model-based reflex agents. These are basically more complex reflex agents with a more advanced decision-making mechanism that can evaluate probable outcomes and consequences before deciding. DigitalOcean writes, "The model tracks how the environment evolves, allowing the agent to infer unobserved aspects of the current state." While these agents don't actually remember past states in the way more advanced agents do, they use their world model to make better decisions about the current state. A use case they give is network monitoring where a model would rely on metrics, logs, events, and network metadata to understand overall network conditions and then, from there, detect anomalies, route alerts, and help with root cause analysis.
Next up, we have goal-based agents or rule-based agents, which, as opposed to reflex agents that act based on rules or world models, goal-based agents can plan sequences of actions to achieve their desired outcomes. The key components of goal-based agents include a goal state, a planning mechanism, a state evaluation, an action selection, and a world model. An example might be an inventory management system that can plan reorder schedules and maintain target stock levels.
Learning agents, as you might imagine, are agentic systems that are capable of improving behavior over time by learning from previous experiences. The key difference from simpler agents is that, rather than relying purely on knowledge that is pre-programmed, they can figure out how to achieve goals through experience. An example might be an advanced customer service chatbot, which doesn't just rely on a set of pre-programmed knowledge, but can interact with and improve its experience over time based on the conversations it's having with customers.
Next up, we have utility-based agents, which differ from goal-based agents in that, rather than shooting for some specific state, utility-based agents can explore and handle trade-offs between competing goals. An example that AWS gives are agents that search for flight tickets that can balance different types of benefits like minimum travel time on the one hand versus price on the other without having to know in advance which of those priorities the end user is going to prioritize most highly.
With hierarchical agents, we start to get into agentic systems. AWS writes, "The higher-level agents deconstruct complex tasks into smaller ones and assigns them to lower-level agents." Each agent runs independently and submits a progress report to its supervising agent. The higher-level agent collects the results and coordinates subordinate agents to ensure they collectively achieve their goals. And the idea here is that, by having different specialized agents that work together in a larger system, you can go out and assign agents to achieve more complex tasks because they can break things down into more manageable subtasks.
And then multi-agent systems might refer more broadly to combinations of these various agents which can achieve more complex goals. So this is a way of breaking down agents based on how they operate in the world. But there is another way to break agents down, which is based on focus.
In this new Information article, they're honing in, it feels to me, on how agents are actually being deployed right now and organizing the categories based on the output that the business who's using them is trying to achieve. The seven categories that The Information lists are business task agents, conversational agents, research agents, analytics agents, developer agents, and domain-specific agents. Business task agents are what some people might huffily say are actually automated workflows. These are useful for fairly simple but repetitive and common use cases like data entry, document classification, invoice processing. A lot of the business process automation layer that's happening right now fits into this category.
The next up is conversational agents, which is inclusive of both external-facing customer service as well as internal-facing support around IT or HR questions. Research agents are, of course, agents that can go do research. Research agents, I think, are particularly important for the average employee because they're one of the first agentic experiences that even non-technical folks are deploying to great effect right now.
The next category that The Information includes are analytics agents, which can analyze structured data to produce graphics, charts, or reports. And then the last two are some of the most discussed categories: developer agents, which, of course, have been the major, major theme for most of this year, and, as we'll see in just a minute, are perhaps the single most significant breakout agent so far. And then, lastly, is domain-specific agents, which have, in other places, been referred to as vertical agents that are specialized agents that have very specific domain knowledge in an area like legal, healthcare, or finance.
So whereas the functional frameworks for understanding different types of agents are useful in understanding what's going on under the hood and how agents are actually interacting with data in the world, these archetypes are a little bit more focused on the types of outcomes that you can enjoy if you successfully deploy these agents, which makes them useful in a different way for a different part of the planning process.
Now, interestingly, we recently got a study from Iconic that looked at how the agent builders themselves are using AI and agents internally. So these are all from firms that are producing AI or agent software in some way, shape, or form. And when it came to the way that they are using AI and agents internally, there are some very clear trends. Notably, coding assistance is by far the most common use case at 77% of organizations. But really, as you can see, there is significant usage of AI and agents across a huge array of business categories from coding assistance to content generation to knowledge retrieval to product design to business intelligence and beyond.
Now, of course, every different agency is going to have some different version of their framework for breaking down different types of agents. KPMG, for example, has tried to simplify that functional breakdown by organizing things into more like four categories as opposed to seven. Their breakdown is the TACO framework that organizes agents into Taskers, Automators, Collaborators, and Orchestrators that are basically divided by the complexity of the tasks that they take on, the amount of human in the loop they need, and the breadth of the systems that they can interact with. What's useful about the TACO framework is that I think that these terms are more intuitive for a lay audience or a non-technical audience than perhaps the breakdown that includes words like reflexivity. Taskers, they write, execute well-defined individual tasks and require a human in the loop. Automators manage more complex tasks that span multi-system workflows. Collaborators are adaptive AI teammates that manage multi-dimensional goals. And Orchestrators are transformative agentic systems that coordinate multiple agents and tools to manage interdependent workflows.
Now, one interesting note and a conversation I'm having a lot at the moment is around this idea of orchestration. It is quite clear, if you're spending any time with enterprises or private equity firms, that there is a huge amount of discussion of orchestrators and multi-agent systems and, hold aside some of the technical conversations around what type of orchestration is needed for multi-agent systems to work, there is very clearly an emphasis on not just individual spot agents but agentic systems. This, to me, was one of the most notable things from Microsoft's Build conference back in May is that, rather than talking about their cool premier agent in each of these different focus categories like conversational agents or research agents, instead Microsoft really put the emphasis on software and agent infrastructure. One of their big announcements was multi-agent orchestration in Copilot Studio that was designed to allow people to deploy more comprehensive and complex agentic systems where the agents could actually interact with one another. We're, of course, also seeing frameworks to support this sort of agent-to-agent interaction including ATA Agent-to-Agent, which is a communications protocol.
And recently someone asked me if I thought that some of these businesses were getting ahead of themselves talking about multi-agent systems when they had barely wrapped their head around or deployed a single agent yet. My short answer was that I thought that no, that was actually a good thing. It's not about skipping steps and not doing the work to actually pilot and deploy and test and learn how to interact with single spot agents. But I think, in general, organizations that think in these systems terms are more directionally aligned with where the world is heading, to the extent that agents really are taking on big chunks of labor and functions that were either (a) previously done by humans or (b) weren't possible to be done by humans because of some complexity or cost equation. They are going to have to work together in comprehensive systems to get the full value of agents. It will not just be a single spot agent deployed in a clever way. It will be big, complex digital worker organizations. And even if we're not all the way there yet, I think anchoring our thinking and our systems design to that agent systems future is going to be more productive than getting lost in the sauce of some specific exciting spot agent.
Now, if that resonates, one recommendation that I would have is to start thinking about the infrastructure and tech stack that's going to need to be put around agents for you to get the most out of them. At the end of that same Iconic report, there's a 12-or-so-page breakdown of different platforms across all of the different agent tool areas. And there are a lot of different agent tool areas: model training and fine-tuning, LLM and AI application development, monitoring and observability, inference optimization, model hosting, model evaluation, data processing and feature engineering, vector databases, synthetic data and data augmentation, coding assistance, DevOps and MLOps, product and design. Now, not every enterprise is going to have to deal with all of these, but some of them are going to be common across basically everyone who's interacting with agents. Inference optimization is going to be something that every organization does; it is nearly 100% the case that you will have some sort of monitoring and observability suite. Same with evaluations. And the point is that, as you think about agent readiness and exploring how to deploy agents, in addition to just thinking about use cases, also think about all of this infrastructure that needs to be built as well.
And, of course, I would be remiss at this point not to point out that, if you are focused on this, if you are on this agentic journey, there are a lot worse places to start than the Super Intelligent Agent Readiness Audit. It's a voice agent that we deploy to interview your leadership and teams about how they work now in order to create a road map and a blueprint for both the specific agentic use cases that are likely to be most valuable for you as well as the additional change management or organizational gaps you need to fill to be able to take advantage of those use cases. Of course, if you are interested, we would love to help you with these problems.
But whether you work with Super Intelligent or not, the reality is that agents are here and they are distinctly not monolithic. They represent a broad set of different types of capabilities and operational models, and understanding and starting to figure out which of those operational models and which of those focus areas are going to be most useful is going to be a key part of your work in the years to come. Hopefully, this gives you some additional tools to think about your options.
And for now, that's going to do it for this AI daily brief. Appreciate you listening or watching as always, and until next time, peace.