Note: Below is are the written notes of above video if you prefer reading:

We are entering a new phase in product design where we create not just tools for users, but AI agents that can act on the user’s behalf.

An AI agent is, in simple terms, an autonomous, goal-driven digital entity that can take actions without constant human direction . Unlike traditional software (which follows hard-coded rules), AI agents exhibit autonomy, adaptability, and learning. They can adjust their actions based on new inputs or changes in their environment and improve from past experience . In other words, instead of strictly scripted interactions, an AI agent can reason and make decisions based on data and goals, not just predefined scenarios . This allows agents to handle unforeseen situations more flexibly than typical programs. For product designers, this shift means designing experiences for delegation: users will state what they want accomplished, and the agent will figure out how to do it.

What Are AI Agents? Fundamental Characteristics

An AI agent is software that uses intelligence to act on a user’s behalf, not just follow fixed instructions. What sets it apart from traditional software are these core traits:

• Autonomy: Once given a goal, it makes decisions and takes action without constant direction — like booking a flight start to finish, not just showing options.

• Reactivity & Proactivity: It responds to changes in its environment (like new data) and can also take initiative to plan ahead.

• Adaptability: It learns from past actions to improve over time — for example, recognizing which emails you usually prioritize.

• Goal-Directed Reasoning: It works toward outcomes, breaking down complex tasks into steps instead of just handling isolated commands.

• Situated Action: It interacts with environments (apps, devices, or even physical spaces), making changes based on what it observes.

In short: AI agents don’t just answer: they act. They carry out user intent through multi-step processes, unlocking new possibilities and design challenges we’ll explore next.

How AI Agents Work: Core Architecture

AI agents think, act, and adapt to accomplish user goals. Unlike traditional systems that react to commands, agents operate in a continuous loop of planning, acting, observing results, and adjusting strategy. To make this possible, an agent combines an AI model with four main components: a reasoning core, a planner, memory, and a set of tools.

1. Agent Core (LLM + Role Prompt)

• At the center is the agent core: usually a large language model (LLM) like GPT-4 or Claude acting as the “brain.” This model is initialized with a system prompt: instructions that define the agent’s purpose, personality, tone, available tools, and behavioral guidelines.

• For example: “You are a travel assistant. Help users plan budget-friendly trips. Use flights and hotels APIs. Be proactive, polite, and optimize for lowest cost.”

• The system prompt gives the agent its identity and tells it how to behave and what it can do.

• Note: You’ll learn how to shape agent behavior using system prompts in Week 3. But just know: the agent’s tone, actions, and logic are often defined in a system message — almost like giving the agent a job description and a personality. More on that soon.

🔧 Behind the Scenes: Model Settings

• The agent’s behavior is further shaped by model parameters like:

  • Temperature: Controls randomness in responses. Lower (e.g. 0.2) = more predictable; higher (e.g. 0.8) = more creative.

  • Max tokens: Limits how much the agent can say or “think” in one go.

  • Top-p / frequency penalties: Fine-tune the diversity and repetitiveness of the output.

• Tuning these lets you calibrate whether the agent feels serious, playful, verbose, or concise — useful when aligning tone with the agent’s role (e.g. legal assistant vs. creative brainstormer).

2. Planning & Reasoning

• Agents don’t just respond, they plan. To achieve a goal, the agent often breaks it down into smaller steps and executes them in order. This can involve:

  • Chain-of-thought prompting: The agent reasons step-by-step to make decisions.

  • Task decomposition: Breaking big goals into sub-tasks (e.g. planning a trip → finding flights → booking hotels).

  • ReAct prompting: A loop where the agent thinks (“I need to check flights”), acts (calls a tool), observes the result, and repeats.

• Some advanced agents even include a critic module — a self-check that reflects on the plan and adjusts if something seems off.
  
  

3. Memory (Short-Term & Long-Term)

• To stay coherent and learn over time, agents rely on memory:

  • Short-term memory: Tracks what’s happening in the current session: like a scratchpad or working memory.

  • Long-term memory: Stores persistent knowledge across sessions, such as user preferences, past conversations, or prior results.

• Example: A personal assistant might remember you prefer window seats and always fly Delta, and use that info automatically in future tasks.
  
  

4. Tools & Actions

• What makes an agent powerful is its ability to act. Tools are external functions or APIs the agent can use — think of them like superpowers. Common tools include:

  • Web search (to gather real-time info)

  • Calendar APIs (to schedule events)

  • Database queries (to retrieve enterprise data)

  • Code interpreters (to run logic or calculations)

• When an agent needs to do something beyond its built-in knowledge, it chooses the appropriate tool and uses it;often via a structured function call.

• Example: If asked, “What’s the weather in Tokyo this weekend?”, the agent may call a weather API instead of guessing: making its response accurate and grounded in real data.

• This tool-based approach allows agents to combine flexibility (from the LLM) with reliability (from deterministic software).

Summary

AI agents combine:

• A reasoning core (the LLM)

• Structured prompting (role, goals, personality)

• Planning logic (step-by-step decisions)

• Memory systems (context and recall)

• Tools (to act on the world)

Together, these pieces allow agents to reason, act, and adapt in real-time — turning AI from a passive assistant into an active collaborator.