Agentic architecture begins with a design decision that is easy to skip: does the problem actually need an agent? A single model call can classify, summarize or generate content. A deterministic workflow can route known steps. An agent becomes useful when the system must choose actions dynamically, use tools, inspect results and adapt its next step based on what happened.<\/p>\n
That distinction sits at the center of Anthropic CCA-F<\/a>. The Claude Certified Architect \u2013 Foundations credential is explicitly concerned with scoping and designing Claude-based solutions, including the choice between agentic and single-shot architectures. Good preparation therefore means understanding why an architecture is agentic, what components make it reliable, and when extra autonomy adds more risk than value.<\/p>\n An agent is not simply a model with a longer prompt. The useful mental model is an LLM augmented with capabilities such as retrieval, tools and memory. Those capabilities let the model obtain information it did not already have, act on external systems and carry forward state that matters to the task.<\/p>\n The architecture should expose those capabilities deliberately. A tool should exist because the model needs an external action or source of truth. Retrieval should exist because relevant knowledge cannot safely fit in the prompt or changes over time. Memory should exist because information must persist beyond the immediate turn.<\/p>\n Adding all three by default usually creates unnecessary complexity. A strong architect starts with the minimum set of capabilities required to complete the task and adds more only when evaluation shows a real gap.<\/p>\n Anthropic’s engineering guidance distinguishes predictable workflows from agents that make more of their own decisions. This is an important architectural boundary. If the sequence of steps is known in advance, a workflow is often easier to test, cheaper to run and safer to operate.<\/p>\n A customer-onboarding process with fixed validation, enrichment and approval steps may be better represented as a workflow. A troubleshooting assistant that must inspect logs, decide what evidence is missing, choose diagnostic tools and revise its hypothesis is more naturally agentic.<\/p>\n The difference is not prestige. Agents are not inherently more advanced in a way that makes them preferable. They trade determinism for flexibility. The architecture is good when that trade is justified by the problem.<\/p>\n A practical agent usually follows a loop: interpret the objective, choose an action, call a tool or produce an intermediate result, inspect the returned evidence, and decide what to do next. The quality of that loop depends on whether the system can obtain ground truth from its environment.<\/p>\n For a coding agent, tests provide evidence. For a support agent, the ticket system and customer account are evidence. For an infrastructure agent, live configuration and deployment results are evidence. If the agent can only reason over its own prior text, errors can compound without a corrective signal.<\/p>\n Architects should therefore identify the feedback mechanism before focusing on prompt wording. An agent that can verify its work is fundamentally easier to trust than one that can only sound confident.<\/p>\n Tools are the interface between a probabilistic model and deterministic systems. Their design is architectural, not cosmetic. A tool should have a clear name, narrow purpose, well-defined inputs and outputs, and predictable failure behavior.<\/p>\n Broad tools create ambiguous choices. A generic \u201cmanage account\u201d tool that can read, edit, delete and approve changes hides too much authority behind one call. Smaller tools such as \u201cget account status,\u201d \u201cupdate billing address\u201d and \u201crequest cancellation\u201d make the agent’s choices easier to evaluate and control.<\/p>\nThe augmented model is the basic building block<\/h2>\n
Separate workflows from agents<\/h2>\n
Design the agent loop around evidence<\/h2>\n
Tool boundaries define what the agent can safely do<\/h2>\n