Designing a Claude Solution: Build Path, Model, Cost and Guardrails

1.Choose how much of the stack to own

Once the tasks and targets are defined, the first architecture decision is who runs the agent loop and the tools. Anthropic describes three paths, which differ "in how much control you keep and how much of the implementation you offload to Anthropic."

The choice also limits where the solution can run. In Anthropic's open-source commerce blueprint, the Messages API and Agent SDK runtimes can run against the Claude API, Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry, or your own gateway. The Managed Agents path runs on the Claude API. A business that is committed to one cloud should weigh that before picking the most managed option.

Sources12

2.Pick a model and tie it to cost

Model choice is where architecture meets the budget. The guidance weighs capabilities, speed and cost, and it describes two ways to start. The cost-first route is to "Begin implementation with Claude Haiku 4.5," test thoroughly, and "Upgrade only if necessary for specific capability gaps." It suits prototyping, tight latency budgets, cost-sensitive deployments and high-volume, straightforward tasks. The capability-first route starts with Claude Opus 5.5 and optimises downward later. It suits complex reasoning, nuanced understanding, and cases where accuracy matters more than cost. The model-comparison guidance notes that most workloads start with Claude Opus 5.5.

Before you switch models, check whether effort can be tuned. Several models support an effort parameter that trades intelligence for latency and cost within one model, and the guidance says "Tuning effort is often a better lever than switching models."

Cost follows volume, so estimate it from the business's real numbers. The moderation guide sizes a platform that receives one billion 100-character posts a month. That comes to about 28.6B input tokens, with 50 output tokens for each of the 3% of posts that get flagged:

Legal summarisation lands the other way. Accuracy has a legal cost, and the whole job is 1,000 agreements, not billions. The guide calls Claude Opus 5 "an excellent choice" there, at an estimated $438.75 compared with $87.75 on Haiku 4.5. What decides the model is the business problem, not a default.

Either way, the decision should be checked against evidence. When you consider upgrading, the guidance says "having a good evaluation set is the most important step in the process." That is the per-task evaluation you defined while scoping.

Sources345

3.Mix models and parallelise to meet cost and time targets

You don't have to run the whole solution on one model. "Multi-model strategies pair a lower-cost model with a frontier model so that most tokens are billed at the lower rate." The guidance names two patterns. In the first, an executor escalates hard decisions to an advisor. In the second, an orchestrator hands bulk work to lower-cost workers. The model table also lists sub-agent tasks among Claude Haiku 4.5's uses. In both patterns, the frontier model's price is paid only where its intelligence is actually needed.

Throughput is part of the design as well. Long documents can take up to a minute each to summarise, so for large collections the legal guide suggests you "send API calls to Claude in parallel so that the summaries can be completed in a reasonable timeframe," within Anthropic's rate limits. The same guide places an ingestion step in front of the model: "Before you begin summarizing documents, you need to prepare your data." That means extracting text from PDFs and cleaning it, and making sure the pipeline can handle every file format the business will actually send.

Sources35

4.Build guardrails into the design

Once an agent can take actions, the design has to decide what it may do on its own. Anthropic's commerce blueprint answers this in three ways. First, facts come from tools. The shopping agent "states products, prices, availability, and store terms only from tool results in the conversation," and cart writes only accept product IDs that a tool returned in that session.

Second, writes are staged. Every change the merchant agent proposes, whether a price move, a restock or a campaign, becomes a staged change with a server-generated ID and a preview card. Guardrails such as maximum price move and campaign budget are checked twice: when the change is staged and again when it is applied. Third, a person approves outside the chat. The approval mechanism differs by build path: a portal button on the Messages API path, a console confirmation prompt in the Agent SDK, or an always-ask permission policy on the apply tool in Managed Agents. The rule is blunt: "An approval typed in chat approves nothing."

The other guides adapt the same idea to their own domain. Moderation can use risk levels instead of a yes/no verdict, because "Creating multiple risk levels allows you to adjust the aggressiveness of your moderation." High-risk posts can be blocked automatically while medium-risk patterns go to human review. Legal summaries should come with disclaimers saying the output is AI-generated and should be reviewed by legal professionals.

Sources24

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.If a Claude solution is too slow or too expensive, the first fix is to switch to a smaller model.Why is that wrong?

   Several models support an effort parameter that trades intelligence for latency and cost within the same model, and Anthropic says tuning it is often the better lever.

   Covered in Pick a model and tie it to cost

2. 2.If the operator confirms a proposed price change in the chat, the agent can apply it.Why is that wrong?

   Staged writes apply only after a person approves them through a channel outside the conversation, such as a portal button, a console prompt or an always-ask permission policy.

   Covered in Build guardrails into the design