Agentic design principles

Who this is for and what you will be able to do

This guide is for teams that are about to design an AI agent and are not sure where to start. By the end you should be able to explain the agentic design principles, apply the three implementation guidelines, and use both to plan a simple agent such as a travel assistant ^[1].

Principles, not architecture

There are many ways to think about building agentic systems, and ambiguity is common in generative AI design. The principles are not a prescribed architecture. They are a starting point for teams defining and building agent experiences ^[1]. They are human-centric UX principles that help developers build customer-centric agentic systems for business needs ^[1].

In general, agents should broaden and scale human capacities such as brainstorming, problem-solving and automation. They should fill knowledge gaps, support collaboration in the ways people prefer to work, and help people become better versions of themselves, for example through coaching on emotional regulation, mindfulness or resilience ^[1].

Agent (Space): where the agent operates

The space principles cover the environment an agent works in, whether physical or digital ^[1]. They break down into three ideas.

• Connecting, not collapsing. Agents connect people with other people, with events and with actionable knowledge. They bring people closer together and are not designed to replace or belittle them ^[1].
• Easily accessible yet occasionally invisible. Most of the time the agent works in the background and nudges only when that is relevant and appropriate. It should be discoverable on any device for authorized users, and it should support multimodal input and output such as sound, voice and text ^[1].
• Visible background work. The agent can move between foreground and background, and between proactive and reactive behaviour, depending on what it senses about user needs. Its background processes and its collaboration with other agents must stay transparent and controllable by the user, even when it is invisible ^[1].

Agent (Time): past, present and future

The time principles describe how an agent behaves across three stages ^[1].

• Past: the agent reflects on history, including both state and context. It gives more relevant results by analysing richer historical data, and it links past events to the current situation ^[1].
• Now: nudging more than notifying. Rather than sending a static notice, the agent can simplify a flow or generate a cue that points the user's attention at the right moment. It delivers information based on context and on user intent ^[1].
• Future: adapting and evolving. The agent adapts to different devices, platforms and modalities, adapts to user behaviour and accessibility needs, and can be freely customized. It is shaped by continuous user interaction ^[1].

Agent (Core): embrace uncertainty, but establish trust

The core principle starts from an honest premise. Some uncertainty in agent behaviour is expected, and it is a key element of agent design. Trust and transparency are the foundation. Humans should be in control of when the agent is switched on or off, and the agent's status should be clearly visible at all times ^[1].

Three guidelines for putting the principles into practice

When you apply the principles, follow three guidelines ^[1].

1. Transparency. Tell the user that AI is involved, how it works (including its past actions), and how to give feedback and change the system ^[1].
2. Control. Let the user customize, set preferences and personalize the system. The user should have control over the system and its attributes, including the ability to make it forget ^[1].
3. Consistency. Aim for consistent, multimodal experiences across devices and endpoints. Use familiar UI and UX elements where possible, and reduce the user's cognitive load with concise responses, visual aids and 'Learn More' content ^[1].

A travel agent example

Suppose you are designing a travel agent. The guidelines translate into concrete product choices ^[1].

• Transparency: say clearly that the travel agent is AI-enabled. Provide start-up instructions, such as a 'Hello' message and sample prompts, and document them on the product page. Show the prompts the user has asked before. Make feedback easy to give, with thumbs up and down or a Send Feedback button. State any usage or topic restrictions openly ^[1].
• Control: make it clear how the user can change the agent after it is created, for example through its system prompt. Let the user choose how verbose the agent is, its writing style and any topics it should avoid. Let the user view and delete associated files, data, prompts and past conversations ^[1].
• Consistency: use standard, recognizable icons. A paperclip means file upload or sharing with the agent, and an image icon means graphics upload ^[1].

Code samples and what to do with this

The sample code is a Python notebook, code_samples/03-python-agent-framework.ipynb, and a .NET file, code_samples/03-dotnet-agent-framework.md, both built on the Agent Framework ^[1]. This article has no install steps. The principles are most useful at the design stage. Before you write a line of code, check the agent's plan against the three guidelines: is it clear that AI is involved, can the user control and forget, and does the interface look and behave the same across devices ^[1].

Adapted from lesson 3 of Microsoft's open-source course *AI Agents for Beginners*^[1], released under the MIT License.