How do I design a complex multi-step workflow?

A well-designed complex workflow is layered: a clear trigger, focused agent steps, deliberate pauses for human input, and explicit paths for when things go wrong.

Start with the trigger and the happy path

Before adding pauses or error handling, map the simplest version:

1. What fires it? (ticket, chat, schedule, webhook)
2. What does the agent need to do? Break it into distinct phases: gather information, act on it, confirm completion.
3. What does success look like? Define a concrete end state.

One agent step per phase

Avoid one enormous agent step that tries to do everything. Instead:

• Step 1 – gather: look up data from HRIS, search the knowledge base, or ask the employee.
• Step 2 – act: draft a document, send a message, create a ticket.
• Step 3 – confirm: post a summary, close the task, or update the original record.

Each step has its own goal, skills, and error path. Smaller steps are easier to debug and easier to change when the underlying process shifts.

Add pauses where humans need to be involved

• Use a ticket reply wait when a step needs input from a specific person already on the ticket.
• Use an approval pause when the agent is about to do something irreversible (sending an email, writing to a system).
• Use a data-gathering sub-agent when you need information from someone outside the current ticket thread.
• Use a scheduled wait when you need time to pass before the next action (for example, waiting 48 hours after a notification before chasing).

Delegate specialised work

If a step requires a back-and-forth conversation with a third party, use delegation rather than loading that complexity onto the main agent:

• Configure the main agent with the overall goal and the skills it needs to orchestrate.
• Add a restricted set of delegatable skills (usually just the data-gathering skill) for the sub-task.
• Write clear instructions for what the sub-agent should gather and what counts as complete.

Example: a full compensation review workflow

Trigger: ticket arrives in the Compensation inbox.

1. Look up employee and policy — agent step, skills: HRIS + knowledge base, no pause.
2. Get manager's rating — agent step with delegation, skill: agentic data gathering, pauses until manager replies.
3. Draft letter — agent step, skill: document generation, no pause.
4. Get HR sign-off — agent step, skill: ticket management, pauses for approval.
5. Send letter and close — agent step, skill: email, no pause.

Guardrails

• Write on error instructions for every agent step. "Post an internal note and assign to HR admin" is a safe default.
• Test the full loop end-to-end in a pilot inbox before enabling on live inboxes. Pay particular attention to the pause/resume paths.
• Avoid chaining too many delegation levels. If you find yourself wanting sub-agents to delegate further, redesign the workflow into two separate workflows instead.
• Keep instructions in each step concise. Long, ambiguous instructions are a leading cause of unexpected agent behavior.