Design a Multi-Step Agentic Workflow

A sales lead asks the system to "set up a follow-up campaign and book time with top prospects." The agent can draft email copy, inspect calendars, and estimate budget in seconds. The risk begins when those drafts become actions. An unconstrained agent might message the wrong recipients, schedule over blocked calendars, or spend budget without review.

This lesson treats that problem as a system design exercise. In this chapter, you move from isolated agent behaviors to governed multistep systems. The design target is not a clever model. It is a workflow that separates planning, approval, and execution so business tasks can complete safely.

We can think of it like a junior operator preparing a packet for a manager. The packet can be generated automatically, but approvals and side effects still pass through explicit gates. That pattern becomes the foundation for the next lesson on customer support agents, where similar controls apply to refunds, escalations, and account changes.

System requirements

Let's turn this goal into concrete functional and nonfunctional requirements.

Functional requirements

• Request intake: The system should parse a natural-language business request into a structured workflow with goals, constraints, and target systems.

• Planning: The system should break the request into steps and select tools within policy constraints.

• Approval gates: The system should pause risky actions, such as sending messages, booking time, or spending budget, until a reviewer approves them.

• Tool execution: The system should execute approved steps through constrained tools and record their results.

• Recovery and audit: The system should resume paused or failed workflows from durable state and keep a full audit trail.

Nonfunctional requirements

• Safety: No side effect should run without passing policy checks and, when risk demands it, human approval.

• Durability: Workflow state should survive crashes and long approval pauses.

• Reliability: Retries should be idempotent, so a resumed step never repeats a side effect.

• Scalability: Planning, execution, and approval-resume capacity should scale independently.

• Auditability: Every plan, approval, and tool call should be traceable for operators and compliance teams.

Resource estimation

Let's estimate the workload, execution capacity, and storage this system needs, starting with the workload.

Workload estimation

Let's start with volume before drawing boxes. Suppose that the platform supports 50 workflows per day. If each workflow averages 6 steps and each step performs 2 tool interactions, the system handles about $50 \times 6 \times 2 = 600$ tool calls per day. If 30 percent ...