In short An agent is a model given not just answering but acting: calling tools, seeing the result and deciding the next step. It runs in a circle — the agent loop: act → observe → decide → again. Plus two amplifiers: tool use (tools) gives the agent hands, extended thinking gives it time to think.
From answering to acting
An ordinary model call (Claude API basics) is "question → answer," one step. But a real task rarely resolves in one step: "find the bug and fix it" needs reading files, running tests, editing code, checking again. To make the model do that, it's put in a loop.
The agent loop
One turn: 1. Gather context (prompt, history, results of past steps) → 2. Action (the model decides to call a tool or give a final answer) → 3. Observation (the tool returns a result, placed back into context) → 4. Decision (task done — exit, or a step needed — repeat). And so on, until the goal is reached.
Analogy: a mechanic fixing a car: looked → turned a bolt → started it, listened → decided, keep going or done. Not one gesture but a loop of trials and checks.
In Claude Platform 101 the loop is first written by hand to understand the mechanics, then "folded up" by the SDK's Tool Runner — it runs the loop for you. In Claude Code 101 the same loop is already built in.
Tool use — tools (the agent's hands)
A tool is a function you describe to the model: its name, what it does, what parameters it takes. The model then decides when to call it and with what arguments; you run the call and return the result to the loop. Without tools the model only talks; with them it acts: searches the web, reads a database, sends mail, runs code. Tool use is what turns a "conversationalist" into a "worker."
Permissions. Since the agent acts, what it may do matters. Permissions bound the available tools and actions — a safety line, especially where the agent changes or deletes things.
Extended thinking — let it think
Extended thinking is a mode where the model is given room to "think out loud" before answering: to unroll its reasoning, break a hard task into steps. It noticeably lifts quality on hard tasks (logic, maths, planning). The reason is the model's nature: it builds an answer step by step (next-token prediction), and explicit reasoning gives it "room" not to jump straight to a conclusion. Analogy: working on scratch paper, not in your head.
How it all adds up
The loop gives structure, tools give capability, thinking gives decision quality, permissions give the line. Then the agent is extended with built-in tools, skills and MCP, and over the long run — managed for context and cost (agent context management). When the whole loop is handed to Anthropic — that's Managed Agents.
Related
The source is Claude Platform 101 and Claude Code 101. Related: Claude API basics, MCP, agent context management, Managed Agents.