What you will be able to do
- Describe the augmented LLM and the client/server tool round trip that underlie every Claude solution
- Tell workflows from agents and pick between them for a given business task
- Match a business problem to prompt chaining, routing, parallelization, orchestrator-workers or evaluator-optimizer
- Break a complex problem into subtasks and decide when to coordinate several agents or models
1.The building block: an augmented LLM calling tools
Every Claude architecture is built from one unit. Anthropic's agent guidance calls it the augmented LLM: a model enhanced with retrieval, tools and memory, which can generate its own search queries, choose tools and decide what to keep. The design advice is to fit these capabilities to your use case and give the model an easy, well-documented interface to them.
With tool use, Claude decides when to call a tool based on the request and the tool's description. It then returns a structured call that your application executes (client tools) or that Anthropic executes (server tools). With a client tool, the response stops with stop_reason "tool_use", your code runs the operation and sends back a tool_result, and Claude uses it to answer. Server tools such as web_search, web_fetch and code_execution run on Anthropic's infrastructure, so you see results without handling execution yourself. When you translate a business requirement, this is the first architectural question for each capability: does it run in your systems, for example against your order database, or is it a capability Anthropic hosts?
Checkpoint 1 of 5· Check yourself
Your solution needs Claude to look up order status in your internal database. How does that tool get executed?
User-defined tools are client tools that run in your application. Claude emits the structured call, and your code executes it and returns the result.
“It then returns a structured call that your application executes (client tools) or that Anthropic executes (server tools).”Source: platform.claude.com
2.Workflows versus agents
Anthropic groups everything built from augmented LLMs as agentic systems, but draws one architectural line through them. Workflows are systems where LLMs and tools are orchestrated through predefined code paths: your code fixes the sequence. Agents are systems where LLMs dynamically direct their own processes and tool usage, so the model decides how the task gets done.
The business translation follows from that line. Workflows offer predictability and consistency for well-defined tasks. Agents are the better option when flexibility and model-driven decision-making are needed at scale. A process the business can already write down as steps, like "draft, check against policy, translate", points to a workflow. A goal whose steps depend on what the model discovers along the way points to an agent. Either one costs more latency and money than a single call, so the step up from a single call has to be justified first.
Checkpoint 2 of 5· Check yourself
A compliance team needs every contract summary to go through the same fixed steps in the same order, and auditors require predictable behaviour. Which design fits best?
Fixed, well-defined steps and a need for predictability call for a workflow. Agents earn their cost when flexibility and model-driven decisions are needed.
“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”Source: www.anthropic.com
3.Five patterns and the problems they solve
Anthropic describes a small set of composable patterns seen in production. Each one answers a different shape of business problem. Prompt chaining exists to trade latency for accuracy by making each call an easier task, and you can add programmatic gates between steps. Routing exists for separation of concerns, because optimizing one prompt for every kind of input can hurt performance on some of them. Ticket routing is the textbook case, and Anthropic's own guide describes it as a type of classification task. Routing can also send easy questions to smaller, cost-efficient models like Claude Haiku 4.5 and hard or unusual ones to more capable models like Claude Sonnet 4.5.
| Pattern | How it works | Use when | Example |
|---|---|---|---|
| Prompt chaining | Each LLM call processes the output of the previous one | The task decomposes cleanly into fixed subtasks | Generate marketing copy, then translate it |
| Routing | Classify an input and direct it to a specialized followup task | Distinct categories are better handled separately and can be classified accurately | Refund requests vs technical support |
| Parallelization | Sectioning (independent subtasks) or voting (same task, several runs), aggregated programmatically | Subtasks can run in parallel for speed, or you need several attempts for confidence | Guardrail screening alongside the core response |
| Orchestrator-workers | A central LLM breaks down tasks, delegates to workers, synthesizes results | You cannot predict the subtasks needed | Coding changes across multiple files |
| Evaluator-optimizer | One call generates, another evaluates and gives feedback in a loop | Clear evaluation criteria and iterative refinement adds measurable value | Literary translation |
Checkpoint 3 of 5· Match them up
Match each business need to the pattern that fits it
Tap a term, then the definition that fits it.
Each pairing follows the source's own examples: voting for vulnerability review, chaining for outline-then-document, routing for customer-service query types, and evaluator-optimizer for literary translation.
“Voting: Running the same task multiple times to get diverse outputs.”Source: www.anthropic.com
4.Breaking down complex problems: orchestrators, subagents and multi-model designs
Task decomposition is how a large business problem becomes something Claude can do well. If the subtasks are known ahead of time, chaining or sectioning hard-codes them. If they depend on the input, as with the number of files a code change touches or the sources a research question needs, an orchestrator breaks the task down, delegates to workers and synthesizes their results.
Claude Managed Agents puts this into practice as multiagent orchestration. A coordinator delegates to other agents, which can act in parallel with their own isolated context. Anthropic says this helps output quality and can shorten time to completion. Each agent runs in its own session thread with its own conversation history, model, system prompt and tools. Tools, MCP servers and context are not shared, though agents do share the sandbox and filesystem. The coordinator fans out independent subtasks and synthesizes the results. Anthropic says multiagent coordination is best suited for complex tasks that span many surfaces, or where multiple well-scoped tasks contribute to one goal. A simple request doesn't need it.
Decomposition is also a cost lever. Multi-model strategies pair a lower-cost model with a frontier model so that most tokens are billed at the lower rate. One common shape is an executor that escalates hard decisions to an advisor. The other is an orchestrator that delegates bulk work to lower-cost workers, which is why Claude Haiku 4.5 is listed for sub-agent tasks.
Checkpoint 4 of 5· Check yourself
In a Managed Agents multiagent session, a subagent needs a fact the coordinator found earlier in its own conversation. What should the designer assume?
Each agent runs in a context-isolated thread with its own history and configuration. Only the sandbox, filesystem and vault credentials are shared, so whatever a subagent needs has to be handed to it.
“Tools, MCP servers, and context are not shared.”Source: platform.claude.com
Checkpoint 5 of 5· Exam question
A company wants a single default model for most of its internal tooling: code generation, data analysis, content drafting, and agentic tool use across many teams, seeking the best combination of frontier intelligence and everyday speed without paying premium enterprise pricing. Which model should they standardize on?
Correct answer: B — Claude Sonnet 5, because it offers frontier intelligence at scale for coding, agents, and everyday workflows at moderate cost.
- A. Incorrect. Opus 4.8 targets complex agentic coding and enterprise work at premium pricing, which is more than most everyday drafting and analysis tasks need.
- B. Correct. Sonnet 5 is described as frontier intelligence at scale built for code generation, data analysis, content creation, and agentic tool use at moderate, introductory pricing.
- C. Incorrect. Haiku 4.5 is a strong economical choice for high-volume or latency-sensitive tasks, but it is not positioned as the best all-around fit for mixed coding, analysis, and drafting work across many teams.
- D. Incorrect. Fable 5 is priced and designed for next-generation, long-running agent intelligence, which is a costlier and narrower fit than the broad, moderate-cost standardization the company wants.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.A more autonomous agent is always the better architecture for a business problem.Why is that wrong?
Agentic systems trade latency and cost for task performance. Well-defined tasks are better served by predictable workflows, or even a single optimized call.
Covered in Workflows versus agents
2.Parallelization and orchestrator-workers are the same pattern because both fan work out to several LLM calls.Why is that wrong?
Parallelization splits work into predefined subtasks. In orchestrator-workers, a central LLM decides the subtasks from the specific input.
Covered in Breaking down complex problems: orchestrators, subagents and multi-model designs
3.Subagents in a multiagent session inherit the coordinator's context and tools.Why is that wrong?
Each agent runs in its own context-isolated thread with its own configuration. Tools, MCP servers and context are not shared.
Covered in Breaking down complex problems: orchestrators, subagents and multi-model designs
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.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 building block: an augmented LLM calling tools“Workflows are systems where LLMs and tools are orchestrated through predefined code paths.”
↩︎ Workflows versus agents“systems where LLMs dynamically direct their own processes and tool usage”
↩︎ Workflows versus agents“The main goal is to trade off latency for higher accuracy, by making each LLM call an easier task.”
↩︎ Five patterns and the problems they solve“Routing classifies an input and directs it to a specialized followup task.”
↩︎ Five patterns and the problems they solve“hard/unusual questions to more capable models like Claude Sonnet 4.5”
↩︎ Five patterns and the problems they solve“This workflow is particularly effective when we have clear evaluation criteria, and when iterative refinement provides measurable value.”
↩︎ Five patterns and the problems they solve“Agentic systems often trade latency and cost for better task performance, and you should consider when this tradeoff makes sense.”
↩︎ Exam trap 1“a central LLM dynamically breaks down tasks, delegates them to worker LLMs, and synthesizes their results”
↩︎ Exam trap 2“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”
↩︎ Checkpoint“Voting: Running the same task multiple times to get diverse outputs.”
↩︎ Checkpoint“but determined by the orchestrator based on the specific input”
↩︎ Prediction - 2.
“Tools differ primarily by where the code executes.”
↩︎ The building block: an augmented LLM calling tools“Your code executes the operation and sends back a tool_result.”
↩︎ The building block: an augmented LLM calling tools“It then returns a structured call that your application executes (client tools) or that Anthropic executes (server tools).”
↩︎ Checkpoint - 3.
“The Claude Agent SDK sits in between, providing the agent loop and tool execution in a process you operate.”
↩︎ Workflows versus agents - 4.
“Ticket routing is a type of classification task.”
↩︎ Five patterns and the problems they solve - 5.
“Agents can act in parallel with their own isolated context, which helps improve output quality and can also improve time to completion.”
↩︎ Breaking down complex problems: orchestrators, subagents and multi-model designs“best suited for complex tasks that either require work across a variety of surfaces, or where multiple well-scoped tasks contribute to an overall goal”
↩︎ Breaking down complex problems: orchestrators, subagents and multi-model designs“Fan out independent subtasks simultaneously (searching multiple sources, analyzing separate files) and have the coordinator synthesize the results.”
↩︎ Breaking down complex problems: orchestrators, subagents and multi-model designs“Tools, MCP servers, and context are not shared.”
↩︎ Exam trap 3“Tools, MCP servers, and context are not shared.”
↩︎ Checkpoint - 6.
“Multi-model strategies pair a lower-cost model with a frontier model so that most tokens are billed at the lower rate.”
↩︎ Breaking down complex problems: orchestrators, subagents and multi-model designs