Prompt & Workflow Design

In an MCP-first system, there is a third building block alongside Tools and Resources: Prompts and Workflows. A workflow is a pre-built sequence that guides an agent step by step through a complex process. It defines which tools are called in which order, where the human must be consulted, and which results are passed on.

Workflows are not scripts. They give the agent context, intent, and structure, without removing its ability to make decisions.

The best UI is often no UI at all

Many tasks in daily work no longer need a dedicated page. They are conversational, contextual, and workflow-based. The human asks, the agent works:

• “Create a summary of all open deals.”
• “Which customers should I contact today?”
• “Schedule a follow-up meeting with all stakeholders.”
• “Prepare an email to this contact.”
• “Check whether any documents are missing for this employee.”
• “Create a report on all projects with risk.”
• “Compare the last two payroll runs.”

No navigation, no filters, no modals. The agent gathers context, combines the right tools, and delivers a result, or asks for approval before acting externally.

Workflows instead of navigation

A classic web app guides the user through navigation: customer list → detail page → communication tab → reminder modal. That is intuitive for humans. For an agent it is unnecessary overhead, it must click through an interface that was not built for the way it works.

A workflow solves the same problem differently: it describes the process directly as an ordered sequence of capabilities. The agent understands what needs to be done, without knowing the surface.

A workflow in detail

The following example shows a buyer follow-up workflow for a real estate software. The agent should check all active projects, identify suitable buyers, and prepare a follow-up communication.

Here is how the workflow runs:

1. projects.list_active, The agent reads all active projects. This is a low-risk read access that can be executed autonomously.
2. contacts.search_buyers, For each project, matching buyer contacts are searched. Also a read access, no external effect.
3. contacts.get_purchase_profile, The agent loads the purchase profile of each contact to check whether a follow-up makes sense.
4. exposes.generate_download_link, For relevant projects, a time-limited download link to the exposé is generated.
5. emails.create_project_draft, The agent creates an email draft with the personalized text and the download link. No sending yet.

At this point the workflow pauses.

The human is asked. The system shows the completed draft: recipients, subject, content, link expiry date, risk level critical. The user can confirm, edit, or reject the draft. Only after explicit approval is emails.send_project_email executed.

6. reminders.create_follow_up, After sending, the agent automatically sets a reminder if the contact does not respond within seven days.
7. deals.recommend_next_action, Finally, the system calculates the recommended next action per deal and returns it as a structured result.

Workflows compose tools

A workflow is fundamentally a composition of individual tools with clear logic:

• Low-risk read accesses (list_active, search_buyers, get_purchase_profile) are executed autonomously and sequentially, no external effect, no risk.
• Write actions with internal effect (create_project_draft, reminders.create_follow_up) are executed but produce no external impact.
• Write actions with external effect (emails.send_project_email) are protected by a confirmation gate. The agent prepares, the human decides.

This pattern applies to every workflow: reading is cheap, writing needs context, external actions need consent. Whoever builds workflows this way keeps the system fully controllable, without paralysing it.