What an agent really is
Not every LLM call is an agent. The distinction matters once you start building seriously.
The typical wrong definition
"An agent is an LLM that can do stuff." That's not enough.
An LLM call is: you put in a prompt, get an answer back. That's a function call, not an agent.
An agent is a system that:
- Pursues a goal (not just answering a question)
- Decides autonomously which steps to take
- Uses tools to interact with the world (read files, search the web, write code)
- Runs in a loop until the goal is reached or given up on
The loop is the difference. An agent observes, decides, acts, observes again. Many times.
Simple example
Not an agent: "Claude, write me an email to customer XY." → Claude writes an email. Done. One turn.
Agent: "Claude, find the 10 biggest customers who haven't ordered in 30 days and send each one a personalized re-engagement email." → Claude has to: read CRM data, filter, sort, look at each customer's history, generate an email, send, write a report. Multiple turns, multiple tools, own planning.
The four anatomical parts of an agent
1. Instructions (the goal + ground rules) "You're a sales assistant. Find customers X. Rule: no customer with a negative note in their timeline."
2. Tools (what it's allowed to use) List of functions with name, description and input/output schema. E.g. get_customers, send_email, log_action.
3. Memory (what it remembers while the task runs) Short-term: what have I done so far. Long-term: what have I learned from earlier tasks.
4. Loop (the execution loop) The agent runner calls the LLM, reads the decision ("I want to call tool X with input Y"), executes it, returns the result to the LLM, asks again. Until the agent reports "done".
Who built that
Claude Agent SDK from Anthropic. OpenAI Agents SDK. Langchain Agents. All three do the same thing: a runner that orchestrates tool calls in a loop.
You can build an agent in 50 lines of code. The complexity difference isn't the loop, it's how well you design the four anatomy parts.
The practical point
Not everything needs an agent. Most tasks can be solved with one LLM call if you build the prompt well. Agents are for tasks with an unknown number of steps.
"Write me this email" → no agent needed. "Research market X and give me a report" → agent. The agent has to decide how deep to dig, which sources, when it's done.
Next lesson
The one pattern 80% of serious agent systems use in 2026: CEO/Worker. One agent that orchestrates other agents. That's Lesson 2.