Agentic Research
中文
Learning roadmap · adapted under MIT

The Agent Learning Roadmap
From your first model call to an agent you can ship safely

This roadmap walks you through three things in order: understand the basics (LLM, prompt, API, and token), then build something (tool calls, agent loops, reading documents, remembering things), and finally make it reliable (permissions, evals, human approval, observability, and failure recovery).

11 stops19 mapped lessonsUpstream project

Two ways through

Track A — CLI Power Userabout 8–10 weeks

Get work done with an off-the-shelf CLI agent: A1 pick a CLI agent → A2 build a repeatable workflow → Stage 5 the ecosystem → A3 plug into a team workflow → Stage 8 interfaces.

Track B — Agent Buildercore line about 16–22 weeks (5–8 h per week)

Build an agent from scratch: start from the first agent loop in Stage 3, walk the Stage 4–8 core line, and finish with the capstone.

The main line: Stage 0 to capstone

Stages 0–2 are the shared base; Track A branches to the CLI route before Stage 3, Track B walks the whole line to the capstone.
  1. 0

    Foundations

    12 mapped lessons

    Call a public API with Python, read JSON, and save your work with Git.

    5–15 hours (1–2 weeks)Open stage
  2. 1

    LLM Basics

    In curation

    See how models go from data to agents, then call an LLM along a repeatable local-to-cloud path; read tokens, context windows, and temperature, and explain model choice with cost and latency.

    5–8 hoursUpstream
  3. 2

    Prompt Design

    In curation

    State goals, data, rules, and output formats clearly, and test the limits of prompting strategies on fixed cases instead of memorizing tricks.

    5–8 hours (with exercises)Upstream
  4. 3

    Tool Use & Your First Agent Loop⭐

    7 mapped lessons

    Build an agent loop with a step limit and argument validation.

    5–8 hoursOpen stage
  5. 4

    Workflow Graphs & Agent Frameworks

    In curation

    Understand the workflow graph first, then build it with a framework; Stage 7 adds evals, observability, approval, and recovery on the same work map.

    10–15 hours (2–3 weeks)Upstream
  6. 5

    Claude Code Ecosystem

    In curation

    The tools-and-rules hub: read the core 5.1–5.4 first and pick from 5.5–5.8 as your work needs; Track A reads how to use them, Track B how to combine them.

    Core line 6–10 h; everything 15–25 hUpstream
  7. 6

    RAG and Memory

    In curation

    Build a minimal RAG and long-term memory flow; store only what is worth keeping, permitted, and deletable.

    6–10 hoursUpstream
  8. 7

    Agent Production Engineering

    In curation

    Like moving a toy car onto a real road: first add the steering wheel, brakes, and dashboard — evals, observability, approval, and recovery.

    12–20 hoursUpstream
  9. 7.5

    Advanced Agentic Choices

    In curation

    It answers one question only: which reproduced failure deserves another layer of checks, fault testing, or planning.

    6–10 hours (selective)Upstream
  10. 8

    Agent Interfaces

    In curation

    Browser use, computer use, and sandboxes: decide between CLI, browser, computer use, or API; start with API/fetch and upgrade only when truly needed.

    6–10 hoursUpstream
  11. ★

    Capstone

    In curation

    Start after A3 or Stage 7: produce a runnable code skeleton, pick the smallest safe interface, and self-check against the rubric.

    3–20 hours (two variants)Upstream

How to learn without getting stuck

  1. 1.Walk one stage at a time: answer that chapter's core question first.
  2. 2.Read the core terms and required reading first: they feed straight into the exercises.
  3. 3.Copy the first command as-is: run the offline test before writing anything from scratch.
  4. 4.Change one thing at a time: rerun the test right after, so you know which change caused the result.
  5. 5.Move on only when you meet the completion checks: understanding it is not the same as doing it.

Adapted from Wenyu Chiou's awesome-agentic-ai-zh (MIT, v2026.09.23, 222 pages): stage order, goals, exercises, and resource tables come from the upstream curriculum, with each stage mapped to this site's existing free lessons. Upstream links checked 2026-08-27. MIT License