What you will be able to do
- Describe the augmented LLM and explain why every agentic pattern is built from it
- Tell a workflow from an agent by asking who controls the next step
- Choose between a single augmented call, a workflow and an agent for a business requirement, using cost, latency and predictability
Key concept
Workflow vs. agent — In a workflow, your code fixes the path that LLM calls and tools follow. In an agent, the model decides its own next step and which tools to use. This one difference drives the cost, latency and predictability of the design.
1.The augmented LLM: the unit every pattern is built from
Every pattern in this subdomain is built from one unit. Anthropic calls it the augmented LLM: a model call extended with retrieval, tools and memory. The model does more than receive these capabilities passively. Current models write their own search queries, pick the tools they need, and decide what information to keep. Chains, routers, orchestrators and autonomous agents are all different ways of wiring augmented calls together. To understand the patterns, first understand what a single call can already do.
On the Claude platform, tools are the main way to add capabilities. Claude decides when to call a tool based on the user's request and the tool's description. That makes tool descriptions part of your architecture, not documentation written afterwards. Anthropic recommends focusing on two things when you build the augmentations. First, tailor them to your specific use case. Second, give the LLM an easy, well-documented interface to them. The Model Context Protocol is one way to do this. It lets a simple client connect to a growing ecosystem of third-party tools.
Checkpoint 1 of 4· Check yourself
A team is adding retrieval and tools to a single Claude call. Which two implementation aspects does Anthropic recommend they focus on?
The guidance names two aspects: fit the augmentations to your use case, and make the interface to them easy and well documented for the model.
“tailoring these capabilities to your specific use case and ensuring they provide an easy, well-documented interface for your LLM”Source: www.anthropic.com
2.Workflows versus agents: who controls the next step
Anthropic uses "agentic systems" as the umbrella term but makes a sharp architectural split inside it. In a workflow, LLM calls and tools follow paths your code defined in advance. The model may produce text or make a classification, but your program decides what happens next. In an agent, the LLM directs its own process and tool use, and controls how the task gets done.
Agents can sound elaborate, but they are usually simple to implement. An agent is typically just an LLM using tools based on feedback from its environment, in a loop. When tools run on your side, the application drives that loop. Claude asks for a tool call, your code runs it and returns the result, and the model chooses its next step. You keep control through stopping conditions, such as a maximum number of iterations, and through checkpoints where the agent can pause for human feedback.
Agents fit open-ended problems where you cannot predict how many steps are needed, and where you trust the model's decisions. Autonomy has a price. Agents cost more, and their errors can compound. Anthropic recommends extensive testing in sandboxed environments, along with appropriate guardrails.
| Dimension | Workflow | Agent |
|---|---|---|
| Who decides the next step | Predefined code paths | The LLM, dynamically |
| Best fit | Well-defined tasks that decompose into known steps | Open-ended problems where the number of steps cannot be predicted |
| Main strength | Predictability and consistency | Flexibility and model-driven decision-making at scale |
| Main risk | Rigid when inputs fall outside the designed paths | Higher costs and compounding errors |
Checkpoint 2 of 4· Check yourself
Which property makes a system an agent rather than a workflow, in Anthropic's sense?
Several LLM calls, tool access or long runtimes can all appear in workflows too. What defines an agent is that the model controls how the task is done.
“Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”Source: www.anthropic.com
Checkpoint 3 of 4· Exam question
A content team needs marketing copy generated and then translated into three languages, with a programmatic check confirming each translated string stays within a fixed UI character limit before it is published. Which architectural pattern best fits this requirement?
Correct answer: A — Prompt chaining, using sequential LLM calls with a programmatic gate that checks translated text length before publishing
- A. Correct. The task decomposes into fixed sequential steps (generate, then translate) with a deterministic, programmatic validation gate between them, which is the defining shape of prompt chaining.
- B. Incorrect. Routing classifies an input to send it down one of several specialized paths; there is no classification decision here, only a fixed sequence of steps.
- C. Incorrect. Orchestrator-workers dynamically determines subtasks at runtime; the languages and steps here are already known and fixed, not discovered dynamically.
- D. Incorrect. Evaluator-optimizer uses a second LLM to iteratively critique and refine output; the scenario describes a deterministic length check, not an LLM-based quality feedback loop.
3.Start simple: match complexity to the requirement
Anthropic's selection rule is simple: find the simplest solution that works, and add complexity only when you need it. Sometimes that means not building an agentic system at all. The reason is economic. Agentic systems often trade latency and cost for better task performance, so each step up in complexity has to pay for itself in the business requirement.
In practice, this gives you a three-step ladder. The first step is a single augmented call, tuned with retrieval and in-context examples. For many applications that is enough. The second step is a workflow. Choose it when a well-defined task needs the predictability and consistency of a fixed path. The third step is an agent. Choose it when the task needs flexibility and model-driven decisions at scale. The platform's multiagent documentation applies the same rule one level up. It reserves multi-agent coordination for complex tasks, such as work that spans many surfaces.
The same caution applies to tooling. Frameworks make it easier to start, but they add layers of abstraction that can hide the underlying prompts and responses, which makes debugging harder. They also make it tempting to add complexity you do not need. Anthropic suggests starting with the LLM APIs directly. If you do use a framework, make sure you understand the code underneath it.
Checkpoint 4 of 4· Check yourself
A company wants Claude to answer employee questions from a fixed HR policy knowledge base. The answers must be accurate and fast. Which starting architecture fits Anthropic's guidance best?
Nothing in the requirement calls for dynamic multi-step control. Anthropic says an optimized single call with retrieval is usually enough, and the more complex options would add latency and cost.
“For many applications, however, optimizing single LLM calls with retrieval and in-context examples is usually enough.”Source: www.anthropic.com
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.Agents are the most advanced pattern, so they are the best default for any production system.Why is that wrong?
Agents and workflows suit different requirements. Workflows give predictability and consistency on well-defined tasks. Agents are justified only when you need flexibility and model-driven decisions, and often a single augmented call is enough.
Covered in Start simple: match complexity to the requirement
2.Any system where an LLM makes a decision, such as classifying an input, is an agent.Why is that wrong?
If the LLM only chooses between paths your code defined in advance, the system is a workflow. It is an agent only when the model directs its own process and tool use.
Covered in Workflows versus agents: who controls the next step
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“Claude determines when to call a tool based on the user's request and the tool's description.”
↩︎ The augmented LLM: the unit every pattern is built from - 2.https://www.anthropic.com/engineering/building-effective-agentsSecondary source
“The basic building block of agentic systems is an LLM enhanced with augmentations such as retrieval, tools, and memory.”
↩︎ The augmented LLM: the unit every pattern is built from“tailoring these capabilities to your specific use case and ensuring they provide an easy, well-documented interface for your LLM”
↩︎ The augmented LLM: the unit every pattern is built from“Workflows are systems where LLMs and tools are orchestrated through predefined code paths.”
↩︎ Workflows versus agents: who controls the next step“They are typically just LLMs using tools based on environmental feedback in a loop.”
↩︎ Workflows versus agents: who controls the next step“The autonomous nature of agents means higher costs, and the potential for compounding errors.”
↩︎ Workflows versus agents: who controls the next step“we recommend finding the simplest solution possible, and only increasing complexity when needed.”
↩︎ Start simple: match complexity to the requirement“Agentic systems often trade latency and cost for better task performance”
↩︎ Start simple: match complexity to the requirement“Workflows are systems where LLMs and tools are orchestrated through predefined code paths.”
↩︎ Key concept“workflows offer predictability and consistency for well-defined tasks, whereas agents are the better option when flexibility and model-driven decision-making are needed at scale”
↩︎ Exam trap 1“Workflows are systems where LLMs and tools are orchestrated through predefined code paths.”
↩︎ Exam trap 2“Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.”
↩︎ Checkpoint“For many applications, however, optimizing single LLM calls with retrieval and in-context examples is usually enough.”
↩︎ Checkpoint - 3.
“Client-executed tools (both user-defined and Anthropic-schema) require your application to drive a loop.”
↩︎ Workflows versus agents: who controls the next step - 4.
“Multiagent coordination is best suited for complex tasks that either require work across a variety of surfaces”
↩︎ Start simple: match complexity to the requirement