色控传媒

By Par Botes, VP AI Infrastructure, 色控传媒

Great technologies give businesses an opportunity to make basic things better. Client/server networking enabled different departments to talk directly without layers of middle management. The Internet very quickly had people making elaborate financial transactions online. Now, agentic artificial intelligence is transforming how computers make decisions on our behalf.

Agentic AI, Explained
The truth is, computers have always been great at automating things, and agentic systems are no different.

Agentic AI isn’t what some have dubbed “an agent,” nor is it a chatbot. It's a system that is allowed to independently take actions to achieve an outcome. Agentic systems adapt to changing circumstances and perform complex tasks, capable of automating multi-step business activities in ways we haven’t seen before. Agentic systems can choose optimized outcomes from several possibilities, developing insights around business operations and customers along the way.

This is why, to many people, agentic AI looks like the kind of magic people experienced when ChatGPT came to the world, only with far more real-world implications. Speaking personally, agentic systems act and deliver complicated insights and outcomes in ways I didn’t initially believe would be possible.

Adopting these systems will come at a cost to traditional processes, just as they did with client/server and the Internet. I believe this will be a net positive: Giving up on current ideas about rigid workflows for computers could ultimately dismantle the legacy of far too many siloed data repositories in enterprise computing.

To many people in enterprise IT, the change is a big and scary one. Workflows are how we currently design and manage computer tasks with precise steps, , and narrowly defined categories of accessible data, among other restrictive things. These are designed to prevent runaway automatons, so saying we should move away from workflows might sound to today’s conventional wisdom like inviting chaos.

It’s not. Workflows are relatively rigid and predefined systems of execution, with the objective of creating predictive repeatability. They were fine under less capable computing systems, drawing from rigid data silos. But these systems are not designed to create an outcome; they were created to automate steps. It’s unclear what in the environment creates outcomes—only what steps, when repeated, deliver a uniform outcome.

Agentic systems, on the other hand, have the flexibility to analyze and understand tasks and dependencies enough to deliver desired outcomes within changing environments. They offer flexibility where there was fixed rigidity, and innovation where there was strict predefinition.

Both workflows and agentic systems are backwards-engineered from a business goal. Agents investigate their environments, evaluating and executing based on what they find via the protocols and interfaces that give it access to data repositories and other agents. Highly capable agents, like the ones we’re starting to see today, can uncover incompatibilities in the system and possibly devise workarounds.

When new information changes the implications of earlier data inputs, they can be designed to make updates without requiring another development and testing cycle. By virtue of examining its environment and commenting on it, agents can add efficiency in addition to carrying out a task.

The feared chaos, whether in the form of bad data “hallucinations” or agents mishandling their tasks, is, of course, a concern. But so were badly designed workflows, or business processes that violated regulations or alienated customers. In every case, the answer is better design and new forms of vigilance.

This brings up another architectural point that is emerging in the agentic AI world. There need to be new forms of and permissions, along with rules that confine the agent. Agents benefit from sub-agents in supervisory, security, and guardrail roles and other dimensions of oversight—not just to prevent bad things from happening, but to keep the ultimate supervisor, the human, in the loop.

In the world of agentic AI, humans are neither idle bystanders nor inventors of the next step the agent takes. People work with agentic systems as ultimate supervisors, arbitrating between approved execution and execution that must be stopped or unrolled. In that sense, the role is not different from other kinds of human management of complex processes.

Even with those restrictions inside the system, agents can deliver extraordinary results. We’ve seen several in our labs, including agents that can build the basics of a platform in minutes that previously took hours for a talented programmer or administrator to set up, test, and validate. One key to our success: adopting a less restrictive approach to siloing data.

This is a policy we’ve had in place for a long time, given the advantages for operational costs, security, and technology-led innovation in areas like machine learning. Data must be stored in more capacious warehouses, lakes, and other less-restrictive repositories.

Building data platforms is commonplace in companies operating public clouds, if only to optimize for agility and efficiency. However, it’s only recently catching on as an architectural pattern for enterprises as data grows exponentially and the means to capture, store, and manage that data have gotten cheaper and easier.

Change has been slow, however. The modern data cloud may be a data best practice, but data is still siloed all over the place in many enterprises. It’s siloed across workload types, departments, and locations, often with no reason other than “that’s the way it’s always been.”

All this to say: The once appropriate data silo is another artifact of legacy technology evolution and now-obsolete technology management paradigms. Different systems had their own data resources, around which jobs and corporate turf grew. Isolation, separation, and rigid rules for configuration of resources followed. In turn, these silos encouraged deterministic thinking, demanding the specific routes and routines of old-style workflows.

That was acceptable for a time, but that time is in the past. In a real-time and highly agile world, with far more abundant and diverse data continually changing and interacting, these practices are a net-negative. As agentic organization becomes increasingly successful and pervasive, it will further diminish these already outdated practices.

This brings me back to my original point, but at a slightly different altitude. New technologies offer a chance to re-examine architectures holistically. They invite us to ask: If we could start fresh with what we know now, would we build our systems the same old way? Often, the answer is “no,” but old habits are only broken by exciting, new technologies. I encourage you to explore agentic systems coupled with modern data and compute platforms to create magic in your data center. I think you’ll be surprised at what’s now possible.

- Par Botes