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

AI agents are transforming how users interact with software: from simply chatting to actually getting things done. In 2025, understanding the types of agents and how they behave is crucial for product designers building the next generation of tools.

🥇 1. Tool-Augmented Agents (Copilots)

The most common and practical agent type in 2025

These agents combine a large language model (LLM) with external tools (APIs, plugins, or functions) to perform real tasks based on natural language input. Instead of just chatting, they act.

🔍 What they do:

• Invoke APIs (like a calendar, CRM, or weather service)

• Run tools like code interpreters, calculators, or web scrapers

• Handle step-by-step task workflows (e.g. summarize → send → file)

🔧 Examples:

• ChatGPT with tools/plugins – can browse the web, run code, analyze files.

• Microsoft 365 Copilot – interacts across Outlook, Excel, and Word to generate content and automate workflows.

• Salesforce Einstein Copilot – uses business logic to answer support tickets or update records.

🎯 Design implications:

• Tool orchestration UX: Users don’t need to see the underlying API, but they should know what the agent did.

• Clear affordances: Let users approve, edit, or override actions taken on their behalf.

• Trust and visibility: Show recent actions, sources, or explanations (e.g., “This summary is based on your last 3 meetings.”)

🛠️ Product pattern:

“Plan my week and reschedule any conflicts.” → Agent checks calendar, finds overlaps, updates meetings, and sends new invites.

🥈 2. Computer-Using Agents (CUAs)

Rapidly emerging — agents that use your UI like a human

CUAs simulate user interaction with the computer itself. They see the screen (via a vision model), understand what’s visible, and control the mouse and keyboard to act — without needing APIs.

🔍 What they do:

• Open and navigate desktop or web apps

• Click buttons, type into forms, scroll pages

• Automate workflows across apps that were never designed for AI
  
  

🔧 Examples:

• OpenAI Operator – a CUA that can control your computer and apps visually.

• Rabbit OS – claims to use CUAs to operate other apps directly.

• Windows Responses API (Microsoft) – agent-driven control over legacy interfaces.
  
  

🎯 Design implications:

• Transparency is essential: Users must see what the agent is doing onscreen.

• Interruptibility: Allow users to pause or take back control at any point.

• Fallback logic: What happens if the screen layout changes or the button isn’t where it used to be?
  
  

🛠️ Product pattern:

“Send an invoice using QuickBooks.” → The agent opens the app, fills in fields, and clicks “Send” — just like a human would.

🥉 3. Autonomous Agents (Goal-Following)

Early-stage but signals the future of AI delegation

These agents don’t just complete one instruction — they take a high-level goal, break it into sub-tasks, reason about steps, and take actions across multiple cycles. They often run in a loop: Think → Act → Observe → Reflect → Repeat.

🔍 What they do:

• Plan multi-step strategies to solve a user goal

• Decide what tools to use and in what order

• Use memory to keep track of progress and refine plans
  
  

🔧 Examples:

• AutoGPT / BabyAGI – early open-source agents that could self-prompt and self-execute tasks.

• Cognosys, TaskMatrix.AI – goal-based frameworks that execute chains of reasoning + actions.

🎯 Design implications:

• Guardrails matter: Without constraints, agents may drift or behave unpredictably.

• Progress UX: Users need to see what’s happening — task status, next steps, success/failure.

• Agent thinking visibility: “Why is it doing this? What’s next?”

🛠️ Product pattern:

“Create a competitive landscape report.” → Agent searches company data, analyzes features, builds a table, generates a slide deck — with minimal user input.

🟡 4. Collaborative & Multi-Agent Systems

Important for more complex, distributed systems

Instead of a single agent doing everything, multiple agents may collaborate — with each one having a specialized role — or coordinate with humans to complete tasks.

🔍 What they do:

• Share data and decisions across agents

• Take on specialized responsibilities in a larger workflow

• Collaborate with users or defer when unsure

🔧 Examples:

• Multi-agent workspaces – e.g., scheduling agent + email agent + note-taking agent.

• Warehouse robots – coordinating to fulfill orders.

• Co-creative tools – like an AI writing assistant that edits while you write.

🎯 Design implications:

• Who owns what? Define agent roles clearly to avoid confusion.

• Coordination UI: Show users how agents are collaborating or escalating.

• Fail-safes: What happens if agents disagree or loop?

🛠️ Product pattern:

“Book my flight and let my boss know.” → One agent handles booking, another drafts the email, a third checks policy compliance.

⚪️ 5. Chat-Only Agents

Still useful, but limited

These agents only engage in dialogue. They don’t remember context, use tools, or act — they just respond.

🔍 What they do:

• Text generation

• Answer questions

• Provide structure for basic conversational UX

🔧 Examples:

• GPT-3.5 in ChatGPT free tier

• FAQ bots, website assistants, simple support chat

🎯 Design implications:

• Best used for constrained, single-turn tasks

• Use clear instructions and context windows

• Avoid giving users the impression the agent can do more than it actually can

🛠️ Product pattern:

“What are your store hours?” → Agent replies with a static answer from a knowledge base.