What you need to know
There are three ways to structure Claude's part in a system. The difference between them is who controls the sequence of steps: your code or the model.
Augmented LLM
One bounded call. You send a request, possibly with retrieval results, tools or extended thinking enabled, and Claude completes one job. Your application decides what happens next; nothing branches on the model's choices. Use it when the task is well defined and the result can be checked.
Example: summarize a customer complaint and draft a reply for an agent to review.
Workflow
The task is split into named steps that your code runs in a fixed order. Some steps call Claude, others call existing systems or apply rules. Because the control flow lives in code, each step can be logged, tested, monitored and audited like any other software. Use it when the steps are known in advance and mistakes are expensive.
Workflows come in four common shapes:
| Shape | How it works | Example |
|---|---|---|
| Chaining | Each step consumes the previous step's output | Extract obligations from a contract, then rate their risk, then write a summary |
| Routing | A classifier picks the path for each input | Billing questions to account data, technical ones to product docs, complaints to a person |
| Parallelization | Independent pieces run at the same time and are combined | Twelve supplier contracts reviewed concurrently, results merged into one report |
| Evaluator-optimizer | One call drafts, another checks against criteria, and the draft is revised until it passes or a retry limit is hit | A customer reply checked against a tone and completeness rubric |
These shapes combine freely. A routed workflow can have a chained branch and an evaluator loop inside it.
Agent
Claude receives a goal and a set of tools and decides for itself which tools to call and in what order. The path is not written down anywhere you can inspect in advance. Use it only when the steps genuinely cannot be listed beforehand, such as investigating an unfamiliar codebase or researching across many sources, and when an unexpected result can be caught and recovered.
A production agent always comes with limits: a minimal tool set, per-turn and total budgets, explicit permissions, a stopping condition, and human approval before consequential actions. These are part of the pattern, not optional extras.
How to choose
Walk through five factors. The first one that rules a pattern out usually decides.
| Factor | Question | Augmented LLM | Workflow | Agent |
|---|---|---|---|---|
| Predictability | Can the steps be listed in advance? | One bounded task | Yes, you wrote them | No, by design |
| Error cost | What does a wrong answer cost? | No guards between steps | Checks between every step | Errors can compound over many turns |
| Observability | Can operations reconstruct what happened? | One call, easy to log | Logs like ordinary code | A long transcript that most tooling cannot alert on |
| Latency | What is the user-visible deadline? | Fastest | Predictable, adds up per step | Open-ended; plan for the worst case |
| Cost | What is the cost per request at volume? | Lowest | Grows with steps | Highest and least predictable |
In the exam
Read the scenario for the deciding constraint. Phrases like "the steps never change", "every decision must be auditable" or "a three-second response" point to a workflow or a single call. Phrases like "the investigation path differs every time" or "unfamiliar codebase" point to an agent, and then the best option is the one that also bounds it with budgets, permissions and approval gates. When two options use the same pattern, prefer the one that adds the controls the scenario's risks call for.
Check your understanding
A legal team wants each contract processed the same way: extract the clauses, rate each clause's risk, then write a summary. Each stage's output must be validated before the next stage uses it. Which pattern fits?
Show answer
Answer: A. The steps are fixed and each one feeds the next, with a check in between. That is a chained workflow. A single call removes the checkpoints, and an agent or multi-agent design adds cost and unpredictability for no gain.
What makes a design an agent rather than an augmented LLM call?
Show answer
Answer: D. Tools and retrieval can be added to a single bounded call. A design becomes an agent when the model controls the sequence of steps toward a goal.
An engineering team needs incidents investigated across logs, dashboards and code, and the path differs every time. Which TWO controls should come with the agent?
Show answer
Answer: A, D. An open-ended investigation suits an agent, but only a bounded one: budgets and a stopping rule cap cost and runaway loops, and approval gates keep consequential actions with people.
Apply it: Choose the pattern for support
Corvane Home wants Claude in its chat and email support. One engineer proposes "a single agent with access to every system". Use the five factors to choose the pattern, then decide how each contact type is handled.
- Score the five factors (predictability, error cost, observability, latency, cost) for Corvane's support, using the case facts.
Done when: each factor has one line that quotes a fact from the case. - Name the deciding factor and the overall pattern it points to.
Why: the exam rewards the simplest pattern that meets the tightest constraint. - Decide the handling for each contact type in the table from the case.
Done when: every contact type has a path, and no state-changing action happens without the right check. - Say what evidence would justify an agent later, if any.
Show model answer
1. Five factors
| Factor | Corvane |
|---|---|
| Predictability | Five known contact types cover all volume; the steps for each are known |
| Error cost | Wrong order status or a wrong refund is costly; refunds over $250 need a supervisor |
| Observability | Support leads must be able to see why a contact was answered as it was |
| Latency | Chat needs a first response within 3 seconds at the 95th percentile |
| Cost | $18,000 a month must cover peak season at 2.5 times normal volume |
2. Deciding factor and pattern. Latency and predictability together rule out an open-ended agent: a multi-turn tool loop cannot reliably answer within 3 seconds, and the paths are known in advance. The pattern is a routing workflow: a fast classifier, then a fixed path per contact type.
3. Handling per contact type
| Contact type | Path |
|---|---|
| Order status | Augmented call: tool call to the Order Management System, then Claude phrases the answer |
| Returns and refunds | Workflow: Claude collects the details, the returns portal applies the rules, refunds over $250 go to a supervisor |
| Product questions | Retrieval over manuals and help centre, answer with a citation |
| Account and billing | Retrieval for policy questions; anything touching payment goes to an agent, and card data never enters the context |
| Complaints | Hand off to a person with a summary of the conversation |
4. Evidence for an agent. Only if production data shows a meaningful share of contacts that the router cannot place and that need several look-ups in an unpredictable order, and only for email, where the 4-hour target leaves room for it.