Software agents have been around for decades. AI in its various permutations has been around for even longer.
Add relative newcomer generative AI (genAI) to the mix, along with increasingly mature approaches to automation, software orchestration, and architecture, and we have all the ingredients we need to build fully mature agentic AI systems.
However, we still lack an understanding of what agentic AI will become when it is fully mature.
Given the potential risks of unbridled AI – Skynet, anyone? – this lack of understanding is dangerous, especially because the core technology is already in place.
All we require to reach this agentic AI end state is the will to do so. We had better understand what we’re getting ourselves in for.
Why People are Afraid of AI Agents
There are several definitions of AI agents available today, but I find the one from Google Cloud to be accurate, comprehensive, and vendor-independent.
According to Google Cloud, AI agents are software systems that use AI to pursue goals and complete tasks on behalf of users. They show reasoning, planning, and memory and have a level of autonomy to make decisions, learn, and adapt. Furthermore, they can converse, reason, learn, and make decisions, and agents can work with other agents to coordinate and perform more complex workflows.
Yet, while AI agents have taken over the hype-o-verse, most agents available today don’t check all the boxes in the definition above. Many of them, for example, can’t learn or coordinate with other agents.
This lack is neither a failing of the definition above nor a mistake on the part of the respective agent providers. Rather, the question of just how much of the definition above applies to the particular agents you are looking at is a question of maturity.
The problem is partly the fact that agentic AI is an emerging technology, and vendors are only now putting together their agentic offerings.
But the larger problem is that people are afraid of AI agents that check all the boxes in the definition above. Autonomous? Make decisions on their own? Adapt? Work with other agents to what purpose? Take over the world, perhaps?
If we had a maturity model for AI agents, say, then we could rank our requirements for the technology from the safer, easier to understand levels up to the most advanced and powerful (and perhaps dangerous) level of maturity.
The good news: there are plenty of agentic AI maturity models out there. The bad news: they generally fall apart at their highest levels of maturity.
What do we want from such models? And how do we get our heads around the concept of mature agentic AI, before it turns into Skynet and takes over the world?
The Elusive Agentic AI Maturity Model
The agentic AI maturity models (AAIMMs, for want of a better name) that I was able to find all more or less agreed on the lower levels of maturity:
- Level 0 (or the precursor to agentic AI) – chatbots, copilots, and other applications of AI that aren’t agentic
- Level 1 – agency at the task level
- Level 2 – simple, deterministic orchestration of agent-driven tasks
- Level 3 – agent-to-agent interactions, generally supporting non-deterministic orchestration
When we get to level 4 and above, however, the various AAIMMs diverge.
The AAIMMs I found don’t provide much clarity at the highest level of maturity. What qualifies as the most mature agentic AI differs from one model to another, often aligning to the value proposition of the company responsible for creating them.
Autonomous behavior and continuous learning only appear in limited ways in the lower levels. Only when we get to the more mature agentic systems do autonomy and learning become central to agents’ behavior.
Here are some examples:
To reach the highest level of maturity on the AAIMM from Vansh Consulting, organizations should focus on human management, ethics, and other human-centric capabilities that a consulting firm might help with.
Jesper Lowgren, chief enterprise architect at DXC Technology, another consulting firm, places the top level of their AAIMM – ‘AI as an autonomous force’ – into the enterprise architecture context for agentic AI.
According to Srini Rangaswamy, MuleSoft solutions engineering leader, his company focuses mature agentic AI on multi-agent orchestration – a particular strength of its platform. Unsurprisingly, MuleSoft’s parent company, Salesforce, aligns its AAIMM with MuleSoft’s.
AgilePoint offers a platform for implementing complex automation scenarios – and such scenarios are at the top of its AAIMM, which it describes as ‘closed-loop automation (self-learning and self-improving).’
From the world of Web3 and blockchain comes Synergetics.ai. According to CEO Raghu Bala, Synergetics.ai’s AAIMM focuses on decentralized agents that leverage blockchain technology.
Perhaps the clearest, most vendor-independent AAIMM comes from Dr. Ali Arsanjani at Google. He places multi-agent systems (MAS) at level 4, reserving level 5 for advanced multi-agent coordination with meta-agents.
MAS feature in many of the AAIMMs. At some level, organizations orchestrate multiple agents, leading to agents that interact with each other.
According to Arsanjani, agentic AI systems eventually feature ‘meta-agents’ whose job it is to coordinate the interactions among agents.
Google’s AAIMM, however, doesn’t stop at five levels. Arsanjani adds a sixth level, adding learning that leverages feedback loops to the model, as mature agents get smarter over time.
This ability for agents to learn and improve on their own and by interacting with other agents who themselves are learning and improving over time is the essential characterization of the most mature state of agentic AI.
From Multi-Agent Systems to Complex Adaptive Agentic Systems
There’s still something missing, however. All the AAIMMs I identified – even Google’s – misses a fundamental fact.
The system that results from large numbers of interacting, self-learning, and self-improving agents is more than a MAS, because its behavior is emergent – that is, the behavior of the system as a whole exceeds the behavior of any of its components.
Systems with such emergent behaviors are complex adaptive systems – and thus the most mature level for agentic AI consists of complex adaptive agentic systems, or what I’ll dub CAAS.
CAAS are self-optimizing, and as such, people cannot directly determine their behavior. Instead, the most that people can do is give them starting points and constraints.
Unlike less mature agents that have specific responsibilities, the starting point for agents in CAAS are descriptions of the business intent for each agent – in other words, a natural language prompt that describes the problem that the human would like the agent to address.
The rest of the ‘programming’ – although programming is the wrong word for self-learning, self-optimizing agents – consists of configuring the constraints to the agents’ behavior. Such constraints also reflect the business intent for the agents individually and the CAAS as a whole.
The infrastructure supporting the CAAS must then be able to translate this business intent into executable technical intent, typically represented as configuration metadata and then continually monitor the behavior of the CAAS to ensure it remains in compliance with the constraints.
What CAAS don’t need is for people to provide anything in the middle between the initial starting point prompt and the intent-based constraints.
Working together, the agents figure out for themselves the best way to achieve the business intent without violating the constraints.
The fundamental point here is that envisioning agentic AI as building increasingly powerful, autonomous agents is missing the central activity of implementing agentic systems, which is building the infrastructure necessary to support intent-based constraints on the behavior of the agents.
Three CAAS Pitfalls
Mature agentic AI systems will be remarkably powerful – perhaps more powerful than any other software that businesses have implemented in the past. And with such power comes concomitant risk.
The most obvious pitfall when implementing learning, self-optimizing agents is to forget that we need constraints. Let the agents improve themselves ad infinitum until – Skynet!
This pitfall, however, is a straw man. Nobody working on agents today is looking to build learning agents that have no controls. Not only would there be hidden risks, but more to the point, such agents would never achieve basic business goals.
In addition to inadequate focus on constraints, a more likely risk is to relegate such constraints to a separate agentic AI governance initiative.
While this approach may be adequate at the lower levels of agentic AI maturity, the constraints are a fundamental part of how to build and run CAAS – so they cannot be an afterthought.
Even with this focus on constraints, one more insidious and very real risk remains.
Static constraints won’t adequately respond to changes in the business intent for the agents. To meet modern business needs, such constraints must be dynamic.
After all, the business intent for the agentic system as a whole will continue to evolve, and the constraints must evolve as well.
It doesn’t take a big leap of imagination to realize that organizations will want to leverage AI to manage and enforce these constraints. After all, translating the business intent for a constraint into its enforceable metadata representation sounds like a perfect task for AI.
Add to the mix that constraints will evolve – and thus managing their configuration drift will also evolve – and leveraging AI is really the only option.
The insidious risk to this entire enterprise is allowing the same AI that runs the agents to manage the constraints.
If there’s any chance that learning, self-optimizing agents can impact the AI that is managing the constraints that must control them, then we’re really asking for the collusion that will inevitably lead to trouble.
The obvious solution to this Skynet-building eventuality is to keep the agents and their constraints separate. Don’t let them interact with each other. Keep the jailers and the prisoners separate, as it were.
Furthermore, the responsibility for keeping these concerns separate should be entirely in human hands. If we let software handle the separation for us, we’re just opening the door for the same problem to rear its head again.
The Intellyx Take
There are several AI agents on the market today that enterprises can purchase and deploy to tackle a wide range of tasks – with many more to come.
The vendors of these agents have built each of them for their respective use cases. Some of the agents have sufficient adaptability to address different tasks, but all of them leave their vendor’s shop with built-in functionality.
As agentic AI matures, this purpose-built agentic nature will give way to general purpose agents that will attempt to accomplish whatever task the user gives them.
They will interact with other agents, call upon diverse data sources, adapt to dynamic situations, and continually learn in furtherance of their user-specified mission.
The constraints will be the only guardrails that can keep such agents on-task.
Crafting, testing, and managing the constraints will be far more than a separate AI governance effort. Instead, such constraint engineering will become the central focus of all agentic AI efforts.
Copyright © Intellyx BV. Intellyx is an industry analysis and advisory firm focused on enterprise digital transformation. Covering every angle of enterprise IT from mainframes to artificial intelligence, our broad focus across technologies allows business executives and IT professionals to connect the dots among disruptive trends. AgilePoint and MuleSoft are former Intellyx customers. None of the other organizations mentioned in this article is an Intellyx customer. No AI was used to write this article. Image credit: Craiyon.