Locate control before naming the architecture

Learning objective

Distinguish augmented LLM, workflow, and agent patterns by who controls the execution path.

Core idea

You already distinguish model-selected actions from prescribed sequences. The architectural addition is to separate the augmented LLM building block from the control pattern around it. An augmented large language model (LLM) has access to retrieval, tools, or memory. A workflow organizes those capabilities through predefined paths. An agent lets the model direct its process as evidence arrives.

These terms describe different aspects of a system. Retrieval does not establish autonomy, and an augmented LLM can be a step inside either a workflow or an agent. Classify the scope being discussed: one answering step, one investigation, or the whole business process.

Anthropic’s Building effective agents uses “agentic systems” as an umbrella while distinguishing workflows from agents by execution control. For an architect, the useful deliverable is a control map: which decisions code fixes, which decisions Claude makes, and which decisions require a person. A product label alone cannot answer those questions.

How it works

Identify what augmentation supplies

Retrieval supplies relevant evidence; tools provide operations; memory supplies retained information. These augmentations address what the model can access. They do not prescribe who chooses the next business stage.

Consider an application that retrieves an approved policy passage before asking Claude to explain it. The application selects the evidence and the answering task. One augmented call may satisfy the need. There is no architectural reason to add an investigation loop when the necessary evidence is already known.

Anthropic’s context-engineering guidance distinguishes prepared retrieval from just-in-time exploration. In the latter, the model uses lightweight references, such as file paths or stored queries, to fetch context as needed. That introduces discretion over evidence gathering. It does not automatically delegate the surrounding approval or delivery process. Record that discretion at the evidence-gathering scope rather than naming the whole system autonomous.

Do not define augmentation as exactly three mandatory integrations or exactly one HTTP round trip. Anthropic names retrieval, tools, and memory as examples of augmentations. A tool-assisted interaction can require further execution, while the architectural distinction remains the owner’s control over the path.

Identify what the workflow fixes

A workflow prescribes the permitted progression. Its steps may contain uncertain model judgments, but its transitions implement known structure. A classifier can choose among predefined handlers without receiving authority to invent a different business process.

Represent the control map with the stage name, its allowed successors, and the owner of the transition. For example, classification produces one of the application’s supported request categories; code selects the corresponding handler. If classification is invalid or unresolved, code selects the exception path. See 1.3.2 for the routing contract.

A workflow can also contain dynamic decomposition. Anthropic places orchestrator-workers among workflow patterns even though the central model determines the subtasks. The surrounding arrangement still delegates work and synthesizes results through the designed structure. Avoid the shortcut “any runtime decision means agent.” Ask which decision is dynamic and how far its authority extends; see 1.3.2 for choosing that workflow shape.

Identify what the agent directs

At an agentic step, the model decides how to pursue the goal from environmental feedback. It can choose further evidence, revise its approach, or encounter a blocker that was not enumerated before execution. The host still defines the environment and permitted effects.

Map those two forms of control separately. Model discretion answers “which investigation should happen next?” Host enforcement answers “which operations can this investigation perform?” An application-operated harness can support autonomous planning, and an SDK-operated harness can sit inside a prescribed workflow. Harness ownership and architectural autonomy are separate axes. In an Agent SDK host, TypeScript Options.tools scopes built-in tool availability, disallowedTools denies tools or matching calls, and a blocking PreToolUse hook enforces per-call policy (see 1.2.2 for its deny output).

For a hybrid, draw the boundary in words: the workflow starts an investigation with a stated goal; the agent selects the investigative actions; the workflow receives its evidence and applies the acceptance gate. If the agent can instead choose whether that gate runs, control over the acceptance process has also been delegated.

Key distinctions and trade-offs

Use this control map to classify a proposed design:

The table does not rank these designs by intelligence. It locates responsibility. A fixed path can still generate an incorrect answer; a flexible path can still have tightly restricted effects.

Prepared context gives the architect a more predictable evidence package. Runtime exploration can discover relevant material that preparation missed, but Anthropic notes that it is slower and can follow dead ends without suitable tools and guidance. Choose discretion where missing evidence changes the task, rather than where a larger tool catalog merely looks useful.

The control map also makes reviews tractable. Review the router’s category contract separately from the investigator’s scope and the final gate. A claim that “the agent handles safety” hides whether safety is a required transition or one possible model-selected action.

Architecture / decision scenario

A government service must answer permit questions within a ten-second interactive target. It also receives appeals whose supporting records differ by case. Staff must authorize every formal decision.

For standard questions, the service retrieves the relevant published rule and asks Claude for an explanation. The architect classifies this answering step as an augmented LLM. The model has evidence, but no authority to choose an appeal process or issue a decision.

For intake, a routing workflow classifies the request into a standard-question handler, an appeal handler, or staff triage. Code owns that finite handler set. The model-selected category does not turn intake into an unrestricted agent.

Inside the appeal handler, a bounded investigator chooses which permitted case records to examine. Findings can expose a missing supporting document and change the next lookup. That step is agentic. Its output returns to mandatory evidence validation and staff authorization before delivery.

The architecture review records all three scopes. Calling the service “an agent” would conceal the interactive path and the approval owner. Calling it “a workflow” without identifying the investigator would conceal where the action sequence varies. The scoped control map explains both latency expectations and responsibility.

Common exam traps

• Classifying retrieval or memory access as proof of an autonomous execution path.
• Assuming every augmented LLM needs all three augmentations.
• Using the number of model calls as the definition of an agent.
• Confusing who runs the harness with who selects the next action.
• Treating dynamic worker assignment as permission to bypass fixed acceptance stages.
• Naming an entire hybrid from the most autonomous step inside it.

What to remember

Augmentation describes access to capabilities. Workflow and agent describe control over progression. Classify the scope, locate runtime discretion, and preserve a separate owner for permissions and acceptance. The resulting control map is more useful than a product’s “agent” label.

Knowledge check

References

• Anthropic, Building effective agents
• Anthropic, Effective context engineering for AI agents
• Anthropic, TypeScript SDK reference
• Anthropic, Control execution with hooks