Prompting academy
  • Level: advanced
  • 10 min read

Agent design: from idea to publication

Building an agent is not a single instruction but a process. Below are five stages of rollout, interface patterns that organise how content is read, and three ways of arranging an agent's work.

This article closes the learning path. It shows how to move from an idea to a working agent, how to arrange the interface it speaks through, and how to organise its work.

FRAME — five stages of building

FRAME organises work on a product into five steps. Each stage ends with something you can check, instead of leaving verification until the end.

  • Formulate — formulate the plan and specification.
  • Render — prepare the visual layer.
  • Attach — connect data and logic.
  • Mock — simulate behaviour and test.
  • Enable — publish under your own address.

Bento Grid — a readable tile layout

Bento Grid is a modular layout of tiles of different sizes, inspired by the Japanese meal box. It shows several pieces of information at once without a feeling of chaos.

  • The most important message takes the largest tile.
  • Side tiles complete the picture.
  • An even grid guides the reader's eye.

Z and F scanning patterns and friction points

The eye moves across a page in a predictable way. Arranging content to follow that movement puts the most important elements where the eye will stop anyway.

  • Z pattern — for simple, short content.
  • F pattern — for longer content and lists.
  • Friction points — places where the reader hesitates: buttons and forms.

Social proof, opinions and ratings belong exactly in those places. Not to persuade by force, but to remove hesitation precisely where it appears.

Animations and transitions without weighing the page down

Smooth scrolling and gentle transitions between views can be achieved with the description language itself, without heavy libraries. The page stays fast while feeling modern.

Rule: an animation should support reading, not distract from it. If it delays contact with the content, it is unnecessary.

ReAct — reasoning and acting

ReAct describes how an autonomous agent works. It plans, performs a step, observes the result and corrects the plan. The loop repeats until the goal is reached.

  • Plan and reflect on the step.
  • Perform the action.
  • Observe the result.
  • Correct the plan and repeat.

Three types of agent logic

The way an agent performs tasks can be described with three patterns. The choice depends on whether order, time or quality matters most.

  • Sequence — tasks run one after another: write, review, correct.
  • Parallel — several tasks start at once to save time.
  • Loop — a task repeats until a condition is met, e.g. until a reviewer accepts the result.

How to combine this into one project

  • Design the agent with FRAME, stage by stage.
  • Arrange content with Bento Grid and the scanning pattern.
  • Place friction points: opinions, ratings, a clear call to action.
  • Write down the working logic: sequence, parallel or loop.
  • Paste the finished prompt into ePersona and test it in conversation.

This is where the Academy path ends. You have a prompt format, a narrative model, a way to research needs, qualification methods and an agent design. Now it is enough to paste the prompt into ePersona and see how it guides a conversation with your customer.

Frequently asked questions

Do I need to know how to code to build an agent?

No. FRAME guides you through the whole process step by step, and the prompt describes the agent's behaviour in language you can understand without code.

Which type of agent logic should I choose?

When order matters — Sequence. When time matters — Parallel. When quality and refinement matter — Loop.

Are animations on a page necessary?

No, but they help. What matters is that they support reading and do not delay access to the content.