Skip to main content
ModelTerms

Learning path · 26 min · intermediate

The agent loop

Tool use, ReAct, plan-and-execute, MCP — how LLMs do work instead of just talking.

An agent is an LLM that can act, not just respond. This path walks from the simplest tool call up through ReAct, multi-agent orchestration, MCP for tool interop, and the modern Computer Use capability that lets models drive desktops.

  1. AgentAI agent

    Why this step: The thing itself — an LLM that picks the next action.

    An AI agent is an LLM-driven system that decides which actions to take, executes them via tools, observes the results, and iterates until a goal is met.

    Read full entry →Agents & Tools · intermediate
  2. Workflow vs Agentworkflow

    Why this step: Anthropic's framing: most production "agents" should be workflows. Set this expectation early.

    A workflow is a deterministic pipeline where humans hard-code the LLM call sequence. An agent lets the LLM decide which steps to take. Anthropic's recommended default is workflow first, agent only when needed.

    Read full entry →Agents & Tools · intermediate
  3. Tool Usetool calling

    Why this step: The mechanism by which an LLM calls external functions.

    Tool use is when an LLM can call external functions — APIs, code interpreters, databases, web fetchers — and read their results. The mechanism that turns chat into action.

    Read full entry →Agents & Tools · beginner
  4. Function Calling

    Why this step: The OpenAI-popularized API shape for tool use.

    Function calling is the specific API mechanism by which an LLM emits a structured request to invoke a named function with typed arguments. The OpenAI-popularized way to do tool use.

    Read full entry →Agents & Tools · beginner
  5. ReActReasoning and Acting

    Why this step: The Thought→Action→Observation loop. The substrate of most agents.

    ReAct is a prompting pattern that interleaves reasoning ("Thought:") with actions ("Action:") and observations ("Observation:"). It is the foundation of most tool-using agents.

    Read full entry →Agents & Tools · intermediate
  6. Plan-and-Executeplanner-executor

    Why this step: The cheaper, more predictable alternative to per-step ReAct.

    Plan-and-execute splits agent loops into a planning step (produce the full step list up front) and an execution step (run each step). Cheaper than per-step ReAct and easier to inspect.

    Read full entry →Agents & Tools · intermediate
  7. Reflexionself-reflection

    Why this step: How an agent learns from its own mistakes within a session.

    Reflexion is a pattern where an agent runs, observes failures, generates a short natural-language "reflection" on what went wrong, and retries with that reflection appended to its prompt — improving via self-critique without weight updates.

    Read full entry →Agents & Tools · advanced
  8. Model Context ProtocolMCP

    Why this step: The open protocol that standardizes tool integration across clients.

    MCP is an open standard for connecting LLMs to external tools and data sources. It defines a JSON-RPC protocol so any client (Claude Desktop, Cursor, IDE plugins) can talk to any MCP server.

    Read full entry →Agents & Tools · intermediate
  9. MCP Server

    Why this step: How you actually expose tools to MCP-compatible clients.

    An MCP server exposes tools, resources, or prompts via the Model Context Protocol so any compliant client (Claude Desktop, Cursor, IDE plugins) can call them without bespoke integration.

    Read full entry →Agents & Tools · intermediate
  10. Multi-Agentmulti-agent system

    Why this step: Coordinating multiple agents. Helpful for hard tasks, dangerous for simple ones.

    A multi-agent system uses several LLM agents that talk to each other — a manager and workers, a debate, a pipeline — instead of a single agent doing everything.

    Read full entry →Agents & Tools · advanced
  11. Agentic Codingcoding agent

    Why this step: The killer app of agentic AI in 2025-2026.

    Agentic coding is an LLM-driven workflow where the model reads code, plans changes, edits files, runs commands, and iterates against feedback — autonomously closing tasks rather than just suggesting code.

    Read full entry →Agents & Tools · intermediate
  12. Computer Useagentic computer use

    Why this step: The newest agent capability — Anthropic/OpenAI models that control a desktop.

    Computer use is an emerging agent capability where the model takes screenshots of a desktop or browser, identifies UI elements, and controls the mouse and keyboard — letting LLMs operate any software like a human would.

    Read full entry →Agents & Tools · advanced

You finished the path.

Now stress-test what you remember.

Take the mixed quiz →Pick another path