How did AI evolve from "chatting" to "doing work"?
About 8 min readTraditional chatbots do only one thing: receive text, output text. You ask a question, it gives an answer. The conversation ends, and nothing actually gets done.
AI Agents are different. They can perceive their environment, make plans, take actions, observe results, and then repeat the cycle. They can use tools, browse the web, read and write files, call APIs — and actually get things done for you.
A chatbot is like a customer service hotline — it can only answer questions. An agent is like a personal assistant — it can book flights, reschedule meetings, and send emails for you.
ReAct (Reasoning + Acting) is the most classic operating pattern for agents. The core idea is simple: the LLM first thinks (Reasoning), then acts (Acting), then observes the result, and repeats.
The full process goes: Thought → Action → Observation → Thought → ... until the task is complete.
Task: "Check tomorrow's weather in Beijing" — click "Next Step" to watch the agent's reasoning process:
call_weather_api(city="Beijing", date="tomorrow")ReAct is like solving a math problem — first think of an approach (Thought), then work it out on paper (Action), then check if it's correct (Observation).
Tool calling (Function/Tool Calling) is the key to an agent's capabilities. An LLM by itself can only generate text, but through tool calling, it can output structured JSON describing which function to call and what parameters to pass.
Important: the LLM doesn't execute tools directly — it only sends requests. An external system handles the actual execution, then returns the result to the LLM to continue reasoning.
Click "Play" to watch how an LLM retrieves information through tool calling:
Tools are like the agent's hands — the brain (LLM) decides what to do, and the hands (tools) carry it out.
Real-world tasks often can't be completed in a single step. For example, "write me a blog post about Apple Silicon" requires the agent to break it down into multiple sub-steps and execute them in order.
The challenges include: error recovery (what if a step fails?), context management (how to filter when there's too much information?), and termination judgment (when is the task actually done?).
Watch how a complex task gets broken into subtasks and executed step by step:
Multi-step planning is like cooking — you don't throw all the ingredients into the pot at once. Instead, you prep, cook, season, and plate in order.
An agent doesn't just chat — it perceives, acts, and observes. At its core is an LLM-driven execution loop.
The Thought → Action → Observation loop lets agents work like humans — thinking and doing, step by step.
Through tool calling, LLMs evolve from text-only generators into all-around assistants that can search, calculate, and execute code.
Faced with multi-step tasks, agents decompose, prioritize, execute sequentially, handle errors, and keep pushing forward.
Agents are AI's evolution from "knowing the answer" to "solving the problem" — not just smarter, but genuinely useful.