The AI System Landscape
Explore the two broad categories of AI systems—LLM-based non-agentic and agentic systems—and understand their architectural differences. Learn how Forward Deployed Engineers evaluate, scope, and hand off these systems by focusing on reasoning models and harness engineering, crucial for ensuring production readiness and operational stability in complex AI deployments.
FDEs encounter a wide range of AI systems across customer environments. Some are straightforward: a model receives an input and produces an output within a single processing cycle. Others operate across multiple steps, calling external tools, maintaining context across a session, and taking autonomous actions toward a goal. These two types of systems are architecturally distinct, and they require different approaches to scope, evaluate, and hand off.
The system type that the FDE recommends during scoping shapes everything that follows. A simpler, non-agentic system has a shorter build cycle, more contained evaluation criteria, and a handoff that the customer’s team can take ownership of quickly. An agentic system introduces coordination complexity, a broader failure surface, and production requirements that go well beyond what a simple pipeline demands. Matching the system type to the problem is one of the most consequential decisions an FDE makes during scoping.
This lesson establishes the landscape: two categories of AI systems, what makes each one distinct, and the structural concept that defines how agentic systems actually work in production.
Two categories of AI systems
AI systems are divided into two broad categories:
LLM-based non-agentic systems
Agentic systems
The distinction lies not in the capability of the underlying model but in how the system operates at runtime.
LLM-based non-agentic systems
An LLM-based non-agentic system takes an input, passes it through a language model, and returns an output. The processing path is fixed. The same input type produces the same sequence of operations every time, and the system does not take autonomous actions based on what it observes mid-process. ...