Do AI Need to “Read the Room” and “Make Plans”? Breaking Down the Three Core Skills of AI Agents
Imagine you hired two assistants.
The first assistant is extremely knowledgeable. When you ask, “How should I plan a five-day autumn foliage trip to Kyoto?” he instantly produces a perfect itinerary, even adding historical anecdotes along the way.But when you say, “Great, now help me book the flights and hotels,” he politely replies:“Sorry, I’m just a language model and cannot access the internet to perform actions for you.”
The second assistant, however, not only gives you advice but also opens a browser, compares prices across hotel booking platforms, checks your calendar availability, pays for the tickets with your card, and finally sends the confirmation email to your inbox.
This illustrates the gap between ChatGPT (Large Language Models, LLMs) and AI Agents.
If an LLM is the “brain” filled with human knowledge, then an AI Agent is a complete organism equipped with eyes, ears, hands, and feet.For AI to evolve from simply “talking” to actually “taking action,” it must master three core capabilities:Planning、Memory、Tool Use & Perception.
Planning: Turning Ambition into Step-by-Step Actions
When humans handle complex tasks, our brains naturally perform task decomposition.
For example, when you decide to cook dinner, your mind automatically breaks it down into smaller steps:
Check the refrigerator > Write a shopping list > Go to the supermarket > Prepare the ingredients > Cook the meal.
Early AI systems often skipped directly to the final answer or produced vague explanations.
AI Agents introduce a crucial planning mechanism, most notably the concept of Chain of Thought (CoT).
- Task Decomposition
An Agent breaks a vague objective into multiple actionable subtasks.
For instance, if you ask an Agent:“Analyze this company’s competitors and write a report.”Its planning module may create a workflow like this:
-
Step A: Identify the company’s core products.
-
Step B: Find companies offering similar products.
-
Step C: Scrape their official websites and financial reports.
-
Step D: Synthesize the information into a SWOT analysis.
-
- Self-Reflection and Correction
A good Agent doesn’t just make plans—it also “reads the room.”In other words, it evaluates the results of its actions.If Step B fails (for example, a website cannot be accessed), the Agent does not simply get stuck. Instead, it triggers a self-reflection mechanism:“Why did this step fail? Should I try another search engine?”This ability for self-criticism and adjustment is essential for agents moving toward autonomy.
Memory: No Longer a “10-Second Goldfish”
Imagine chatting with the same AI every day, but it forgets who you are each time. That would be frustrating.
Traditional LLMs are limited by their context window. When conversations become too long, earlier information disappears.
AI Agents address this by mimicking human cognition with two types of memory.
- Short-Term Memory
This is equivalent to human working memory.
The agent records:Details of the current task、Newly obtained data、The ongoing conversation state.This is usually implemented through prompt engineering, where relevant information is temporarily included in the model input.
- Long-Term Memory
This is where Agents truly become powerful.Through Vector Databases, an Agent can store accumulated experience, user preferences, or even entire manuals in an external memory system.
When information is needed, the agent uses Retrieval-Augmented Generation (RAG) to retrieve the most relevant knowledge.It’s like the Agent carrying a personal notebook, able to flip through last year’s conversations or professional references anytime.
Tool Use & Perception: The Hands and Eyes of AI
This is the most exciting part of AI Agents.
An agent that only thinks is a philosopher.
An agent that can use tools becomes a productivity engine.
- Tool Use / API Integration
Through APIs (Application Programming Interfaces), an agent can interact with digital tools.
Examples include:
- Calculator – compensating for LLM weaknesses in precise math
- Search engines – retrieving real-time information instead of relying solely on training data
- Code execution environments – writing and running Python scripts for data analysis
- Hardware control – connecting to smart home devices to turn on lights or adjust temperature
- Perception
What we metaphorically call “reading the room” becomes environmental awareness for AI Agents.
When an agent performs an action—such as clicking a web button—it observes what happens next:
"Did the page change?"、"Did an error message appear?"
This multimodal perception allows agents to interpret:images、videos、audio.
As a result, they can respond in more human-like and context-aware ways.
The Future Challenge: When Agents Become Autonomous
Once AI Agents possess planning, memory, and execution capabilities, they evolve into Autonomous Agents.
While this brings enormous convenience, it also raises important questions.
Guardrails
Imagine an Agent trying to achieve the goal “save me money” by canceling all your subscriptions without asking you.Clearly, that’s not what we want.We need guardrails—clear ethical and logical boundaries—to ensure agents pursue goals without violating human values.
A New Era of Human–AI Collaboration
In the future, we may see Multi-Agent Systems.Imagine this scenario:Your personal agent automatically negotiates with an airline’s customer service agent to reschedule your flight.
You only need to press “Confirm.”
Conclusion: The Cambrian Explosion of AI
AI is transitioning from simple question-and-answer systems to action-oriented agents with core capabilities.
This marks the beginning of a new stage in artificial intelligence.
AI is no longer just chatting with us through a screen.
It is entering our digital lives—learning to think, plan, and act like humans.
Learning to “read the room” enables better collaboration.
Learning to “make plans” enables greater efficiency.
As these three capabilities continue to mature, AI will no longer be just a tool—it will become our most reliable digital partner.
