Reasoning & Planning: If you say "Organize a 3-city trip for me," the agent doesn't just list cities. It creates a plan: 1. Check flights, 2. Compare hotels, 3. Cross-reference with my calendar.
Tool Use (Function Calling): The agent can actually "click" things. It has access to APIs, web browsers, and code execution environments. It can check your bank balance, send an email, or run a Python script to graph data.
Memory (Short-term & Long-term): It remembers what happened in step 1 of a task while it's on step 50. It can also store information in a "vector database" to remember your preferences across different days or sessions.
Self-Correction (The "Reflection" Loop): If an agent tries to log into a site and fails, it doesn't just give up. It looks at the error message, realizes it used the wrong password format, and tries a different approach.
Perceive: Analyzes the user's goal and the current environment.
Plan: Breaks the goal into a sequence of sub-tasks.
Act: Executes a sub-task (e.g., searches Google, writes a file).
Observe: Looks at the result of the action. Did it work?
Iterate: Refines the plan based on the results and repeats until the goal is met.
Software Engineering: Tools like Devin or OpenDevin. You don't just ask them for a snippet of code; you give them a GitHub issue link. The agent clones the repo, finds the bug, writes a test, fixes the code, and submits a Pull Request.
Personal Assistants: A "Personal AI Agent" that has permission to manage your email and calendar. It doesn't just "remind" you of a meeting; it sees a conflict, emails the other person to reschedule, and finds a time that works for both of you.
Robotics: As we discussed with Unitree, the "brain" is now agentic. A robot isn't just following a path; it's "trying to clean the kitchen," which involves identifying trash, navigating obstacles, and deciding which cleaning tool is appropriate.