The Autonomy Spectrum: Augmented LLM, Workflow, Agent (CCAR-P)

One spectrum, three patterns

The three core architectural patterns, augmented LLM, workflow, and agent, are easiest to understand not as separate boxes but as points on a single spectrum. The Claude Certified Architect - Professional (CCAR-P) exam treats placing them on that spectrum as a remember-level skill because the axis, how much control-flow authority is handed to the model, is what every later pattern-selection decision turns on. At one end, the surrounding code holds nearly all the control; at the other, the model decides its own next step. Knowing where each pattern sits is the vocabulary you reason with when a problem needs more or less autonomy.

An augmented LLM is a single call enhanced with tools, retrieval, or structured output: Claude assists one bounded step and the code stays in charge of everything around it. A workflow is a fixed, predetermined sequence of calls, chaining, routing, parallelization, or evaluator-optimizer, where humans design the control flow and the model fills in each step. An agent is where Claude plans and executes its own sequence of steps toward a goal, deciding what happens next until a stop condition. Autonomy rises left to right; control retained by code falls.

The augmented LLM: one enhanced step

The augmented LLM is the lowest-autonomy pattern: a single call that has been given more to work with, tools it can invoke, retrieved context, a structured output schema, but still a single bounded step. The code decides when to make the call and what to do with the result; the model does not decide what happens next. Importantly, "augmented" does not mean "no tools." An augmented LLM can call a tool within its one bounded step; what makes it augmented rather than agentic is that it does not run an open-ended loop choosing its own sequence.

The workflow: humans design the sequence

A workflow sits in the middle. It involves multiple Claude calls, but the sequence is fixed and designed by humans: the control flow is predetermined, whether that is a chain where each output feeds the next, a router that classifies and dispatches, a parallelization that fans work out and aggregates, or an evaluator-optimizer that generates and checks. The model executes each step, but it does not decide the order or choose to add steps. The authority over control flow stays with the human designer, which is exactly what distinguishes a workflow from an agent.

The agent: the model holds control flow

An agent is the high-autonomy end. Claude is given a goal and the means to pursue it, and it plans and executes its own sequence, choosing each next step based on what it has learned so far, until a stop condition fires. The control-flow authority has been handed to the model. This is powerful for open-ended tasks whose steps cannot be predetermined, and it is exactly why agents are harder to observe and more expensive: the sequence is decided at runtime by the model, not written in advance by a person.

In production, an agent is never handed unbounded autonomy. It is fenced by four constraints the course treats as mandatory, not optional: a constrained tool entry point (the model can only reach the tools it is given), a per-turn budget, explicit permissions on what those tools may do, and stopping criteria that end the run. These bounds are what keep an agent from becoming a liability, and their absence, tools that change state with no turn limit and no way to measure whether the goal was met, is the canonical way an agent design goes wrong.

Two axes underneath the one spectrum

The single spectrum is a compression of two independent axes the module plots each pattern on: how predictable the path through the work is, and how much autonomy you are willing to grant the model. An augmented LLM sits in the high-predictability, low-autonomy quadrant, a workflow occupies the predictable-shape middle band where model judgment is bounded inside each step, and an agent sits in the low-predictability, high-autonomy corner where the model owns the trajectory. Reading a design on both axes, rather than on a single dial, is what stops you reaching for an agent when the path is actually enumerable and only the authority felt uncertain.

What the CCAR-P exam trips candidates on

The exam tests two traps. The first is treating workflow and agent as synonyms because both involve multiple Claude calls. The number of calls is not the distinction, control-flow authority is. A workflow's sequence is fixed by a human; an agent's is decided by the model. A scenario that describes multiple calls does not by itself indicate an agent, and the credited answer looks at who decides the sequence.

The second is assuming an augmented LLM means no tool use. An augmented LLM can call tools within a single bounded step; what makes it augmented rather than agentic is the absence of an open-ended loop. A scenario may describe a single tool-using call and tempt you to label it an agent, when the reliable reading recognises it as an augmented LLM because the model is not choosing its own multi-step sequence.

Common misreadings to avoid

How this shows up on the exam

Domain 1 questions on this knowledge point describe a design and ask you to classify it, or ask what distinguishes the three patterns. The reliable reading places each on the autonomy spectrum by asking who holds control-flow authority: code for an augmented LLM, a human-designed sequence for a workflow, and the model itself for an agent, regardless of call count or whether a single call uses a tool.

This spectrum is the foundation for the rest of the pattern-selection task. It expands into the workflow sub-patterns, grounds the five deciding factors that choose among the three, and underlies orchestrator-subagent fan-out and fan-in. It also frames recognising non-determinism as a feature vs a defect, which is sharpest at the agent end.