Agentic AI is complex, not complicated

There’s a lot of interest in and concern around the use of AI agents. For organizations grappling with whether and how to use agentic AI, I recommend considering the model from the perspective of complex—rather than complicated—systems. Indeed, accepting the fact that agentic AI is complex rather than complicated will be key to harnessing its power and applying necessary protections and controls.

What’s the difference between complex and complicated? Computer science, for example, involves complicated systems—relating to cause and effect from an engineering perspective. Anthropology, on the other hand, involves complex systems—where you can’t control every variable and you have to focus instead on “factors,” as they call them in finance.

In complex systems, we have confidence intervals about what we think is happening. We can be, for example, 60% sure or 85% sure, but we can never be absolutely sure. Often, we can get to the right answer for the wrong reasons. We can even get to the wrong answer for the right reasons for any outcome below our confidence interval. Outcomes are innately multi-variate, and it’s impossible to know why they turned out the way they did.

Here are some examples of complicated and complex systems that my technical peers—programmers, systems administrators, architects—will likely connect with:

Writing Python code is complicated; managing Python programmers is complex.

Editing a video is complicated; making a video go viral on YouTube is complex.

Compiling a C program is complicated; doing a YOLO run when training a base model is complex.

DNS lookups are complicated; running a registrar is complex.

Registering CVEs is complicated; predicting how a hacker will use a CVE is complex.

Now let’s apply the model to autonomous agents. Redesigning your automation infrastructure is complicated; letting an AI agent commit new code with no human intervention is complex—even scary. But there are techniques that we can use to reap the many benefits of agentic AI while acknowledging and addressing its complexity. For example:

Think statistical: The outcomes in our lives feel deterministic, but they’re not. When you back up and analyze human populations at scale, our decisions are statistical in nature, such as how a certain percentage of people will vote one way or another in an election. The process large language models (LLMs) use to drive agents is also statistical in nature, but the outcomes are less precise than they would be with humans, so you have to check the work—or, better yet, write another agent to check the work for you. (Yes, that can work.)

Focus on factors: Financial markets are a complex system, driven by unpredictable fluctuations, so the focus is on factors—the forces that have historically driven asset returns—instead of individual fluctuations. There should be a similar focus in software systems. For example, we can create agents for a senior engineer to do architecture, a junior engineer when we don’t want to change the architecture, a quality engineer to keep them both honest, and an auditor to check all of them to make sure they’re not colluding. We understand the factors of software production and what driving forces each of those factors contribute to the system. You can’t predict what each individual actor will do, but when they each have a specific job, their forces will act in concert to create better software.

Use heuristics and signals: In systems biology, there is no way to model every interaction in the system, so we statistically predict what’s likely to happen. We do that a bunch of times in a row, then analyze statistically what the most likely outcome is, to increase confidence. As the system becomes more statistical in nature, so too does the testing framework need to adapt. We already do this today with organizational security training. We know that some people will make a mistake when faced with social engineering. We can improve the chances that people recognize the attack, and resist it, but we cannot completely remove the risk. Kubernetes is another good example. We run multiple pods because we know that some might fail. We have to build these same kinds of heuristics into agentic AI processes.

Do digital-to-analog conversions: We do this with audio signals all of the time and don’t think about it, but it’s also common with other problem sets like sequencing DNA. (Polymerase chain reaction is a good example.) If you can’t track every discreet state, listen for signals. For agents, this means using digital approval processes, ticket systems, etc., to ensure that agents interact. This will create discreet states where information flows between agents. This model also has the bonus of creating separation of powers and letting the agents “cheat” when performing tasks such as committing code and fixing problems.

Switch between deterministic and statistical models: In a deterministic world, things flow logically. Complex systems like LLMs are non-deterministic. Agents glue these two worlds together with Model Context Protocol (MCP) servers. The more work that gets done in this deterministic world, the more you can trust the results. For example, an agent will gather accurate information and context by accessing file systems and databases or by running commands (governed through MCP). With that said, while agentic AI is statistical in nature, AI practitioners should lean into deterministic tools and APIs when possible.

Understanding the difference between the merely complicated (deterministic world) and truly complex (non-deterministic, statistical world) is key to thriving in this new world of AI in general and agentic AI in particular.

Complex systems are exhilarating because they’re so unpredictable. When it comes to agentic AI, we need to be open to the fact that the technology will turn many of our complicated systems into truly complex systems. Effectively managing these complex systems will be an important part of our job in this new world.

—

New Tech Forum provides a venue for technology leaders—including vendors and other outside contributors—to explore and discuss emerging enterprise technology in unprecedented depth and breadth. The selection is subjective, based on our pick of the technologies we believe to be important and of greatest interest to InfoWorld readers. InfoWorld does not accept marketing collateral for publication and reserves the right to edit all contributed content. Send all inquiries to doug_dineley@foundryco.com.