Designing an End-to-End Agent Workflow
Explore how to design Claude-based agent workflows starting from ambiguous customer requests to detailed system designs. Understand how to clarify problems, select workflows or agents, map data and tools, build safety features, and plan evaluation and deployment. Gain skills to create clear, reviewable designs that align technology with customer needs before coding begins.
Customers rarely arrive with a specification. A procurement director says the team wants AI to handle purchase order approvals, and that sentence is the whole brief. Behind it sit questions nobody has answered yet, such as which orders count, who approves today, what systems hold the data, and what happens when the AI is wrong.
An FDE's first deliverable in that conversation is a design. It names the parts of the system, the decisions behind each part, and the risks the design accepts. The same skill is what interviewers test when they ask a candidate to design a Claude system live, often in under an hour, for a problem described in a few sentences.
A design built through a fixed method holds up better than one improvised under pressure. The method below works for most Claude systems, and it builds safety in as part of the design.
A method for designing Claude systems
The method moves from the problem toward the technology, and each step depends on the answers before it.
Clarify the problem: Who uses the system, what decision or task it supports, how much volume it handles, and what success looks like.
Choose the shape: Decide whether the task needs a fixed workflow, an agent, or a mix of both.
Map data and tools: List the systems the design touches, separating reads from actions that change records.
Choose models and context: Pick a model for each step and decide what each request must contain.
Build in safety: Set permissions, human approval points, and boundaries for untrusted content.
...