AI as a System (Part 5): Agents - When AI Starts Doing Instead of Answering

This is Part 5 of the AI as a System series.
See the full series here.

The Shift from Execution to Systems

In the previous layer, execution introduced the ability for AI to take action within an environment. The system could read files, run commands, and produce results that extended beyond a single response.

The next layer builds on that capability. Instead of only taking actions when directed, the system begins determining which actions to take and how to adjust over time.

This is where agents emerge.

What an Agent Actually Is

An agent is not a specific product or interface. It is a pattern that emerges when reasoning, access, and iteration are combined into a system.

An agent is a system that can take a goal, determine how to achieve it, and take actions until it reaches an outcome.

This changes the role of the system. Instead of responding to a single request, it works toward a result across multiple steps.

The Role of Iteration

The key difference between earlier layers and agents is iteration.

In earlier interactions, the process is linear. You provide input, the system responds, and you decide what happens next. Each step requires a new prompt or instruction.

With an agent, the system continues operating after the initial request. It evaluates results, adjusts its approach, and proceeds toward the goal without requiring direction at each step.

This introduces a loop where the system generates output, takes action, observes the result, and refines its next step based on what it learns.

What Agents Are Built From

Agents do not replace the earlier layers. They depend on them.

At the core is a model that provides reasoning. That reasoning is connected to tools and an environment that define what actions are possible and where those actions occur. A control loop manages how the system moves from one step to the next, using results from previous actions to inform future decisions.

Each of these components is necessary. Without reasoning, the system cannot decide what to do. Without access to tools and an environment, it cannot act. Without iteration, it cannot make progress over time.

What Agents Can Do

When these components are combined, the system is able to handle tasks that extend beyond a single interaction.

It can generate and modify code, run commands, query APIs, read and write files, execute workflows, and respond to errors as they occur. More importantly, it can coordinate these actions over time in pursuit of a goal.

This coordination is what distinguishes agents from earlier layers. The system is not just acting. It is acting with direction and continuity.

Why Agents Feel Different

The experience of using an agent is noticeably different from earlier interactions.

Instead of directing each step, you define a goal and observe how the system works toward it. The role shifts from issuing instructions to evaluating outcomes and guiding the process when needed.

This changes how work is structured. The system becomes part of the workflow rather than a tool used at individual steps.

The Tradeoffs

Agents introduce new capabilities, but they also introduce new challenges.

They can take inefficient paths, misinterpret goals, or introduce unintended changes if not properly constrained. Because they operate over multiple steps, small issues can compound if they are not detected early.

For that reason, most implementations include safeguards such as scoped permissions, testing, and human review. The system becomes more capable, but it also requires more deliberate oversight.

The Beginning of System-Level Thinking

Agents represent a shift from using AI as a helper to incorporating it as part of a system.

Instead of focusing on individual interactions, the focus moves to defining goals, constraints, and boundaries. The system is responsible for execution and iteration within that structure.

This shift has implications for how systems are designed and how work is coordinated. Responsibility begins to move from manual processes to system-driven workflows.

What’s Next

At this point, the system can generate, act, and iterate.

The next layer focuses on how AI connects to real data, enabling it to operate with information that is specific to your systems rather than relying only on general knowledge.