Learning Path

Seven tracks, 45 steps. Start at the top if you are new. Jump to any track if you already have a working setup.

PATH COMPLETION PROGRESS0 / 45 STEPS (0%)
TRACK 016 Steps

Agent Foundations

What an agent actually is, what a harness does, and getting a working setup on your machine.

  1. 01
    What agents and harnesses actually are
    The difference between a model, an agent, and a harness, explained without jargon, so the rest of the ecosystem makes sense.
    7 min
  2. 02
    Picking your first harness
    An honest comparison of the main agent harnesses and how to choose one, based on how you already work rather than on benchmarks.
    8 min
  3. 03
    Your first working setup
    Get an agent installed, connected, and doing real work on your own project, with the mistakes that trip up most people called out.
    10 min
  4. 04
    The agent loop
    The plan, act, observe cycle every coding agent runs, and why understanding it makes agent behaviour predictable.
    7 min
  5. 05
    Permissions and approvals
    How to decide what an agent may do without asking, what needs a prompt, and what it should never touch.
    8 min
  6. 06
    When not to use an agent
    The tasks where an agent costs more time than it saves, and the honest signals that you should do it yourself.
    6 min
TRACK 027 Steps

Prompting and Skills

Writing instructions an agent follows reliably, and packaging them as reusable skills.

  1. 01
    How models read instructions
    Why some instructions get followed and others get ignored, and what to change when an agent keeps doing the wrong thing.
    7 min
  2. 02
    Writing a project rules file
    What belongs in AGENTS.md or CLAUDE.md, what does not, and why short files outperform thorough ones.
    8 min
  3. 03
    Skills explained, and installing your first
    What a skill is, how the SKILL.md format works, when a skill beats a prompt, and how to write one that actually gets used.
    8 min
  4. 04
    Writing your first skill
    Build a working SKILL.md from scratch, and learn the description field that decides whether it ever loads.
    9 min
  5. 05
    Prompt patterns that work
    Five reusable prompt shapes for agent work, with the situation each one is for.
    8 min
  6. 06
    Common prompting mistakes
    The mistakes that waste the most time, why each one fails, and what to do instead.
    7 min
  7. 07
    Subagents and delegation
    When to split work into a separate agent with its own context, and when a subagent makes things worse.
    8 min
TRACK 036 Steps

Context Engineering

Deciding what goes in the context window and what stays out, which is most of the job.

  1. 01
    What is the context window
    What actually occupies the context window, why it fills faster than you expect, and what happens when it runs out.
    7 min
  2. 02
    Context engineering basics
    Why agents degrade on long tasks, what actually fills a context window, and the habits that keep output quality high.
    8 min
  3. 03
    Context rot and why more is worse
    Why filling a large context window degrades answers, and how to keep the signal to noise ratio high.
    7 min
  4. 04
    Compaction and summarisation
    How agents keep working past the context limit, what gets lost when they do, and how to control it.
    7 min
  5. 05
    Managing tool output
    Tool results are usually the biggest thing in your context. How to stop them crowding out the work.
    7 min
  6. 06
    Working across multiple sessions
    How to carry state between sessions so a new one starts productive instead of confused.
    7 min
TRACK 047 Steps

Tools and MCP

Giving an agent real capability through tool calls and Model Context Protocol servers.

  1. 01
    What tool calling actually is
    The mechanism behind every agent action, and why the model never runs anything itself.
    7 min
  2. 02
    MCP explained, and connecting your first server
    What the Model Context Protocol is, how to connect a server to your agent, and the security decisions you are making when you do.
    9 min
  3. 03
    Installing your first MCP server
    Get the filesystem server running, understand the config format, and know how to tell whether it worked.
    8 min
  4. 04
    Choosing MCP servers safely
    An MCP server runs with your permissions on your machine. How to judge one before you install it.
    8 min
  5. 05
    Building a simple MCP server
    Write a working MCP server that exposes one tool, and connect it to your harness.
    10 min
  6. 06
    Tool design that agents can use
    The design choices that decide whether an agent uses your tool correctly or fumbles it every time.
    8 min
  7. 07
    Debugging tool calls
    A systematic way to find out why an agent is not calling your tool, or calling it wrong.
    7 min
TRACK 056 Steps

Retrieval and Memory

Retrieval augmented generation, embeddings, and how agents remember across turns and sessions.

  1. 01
    Why retrieval exists
    The problem RAG actually solves, and why the obvious alternatives do not work.
    7 min
  2. 02
    Embeddings explained
    How meaning becomes numbers, why that enables search without keywords, and where it breaks down.
    8 min
  3. 03
    Chunking and indexing
    How you split documents decides retrieval quality more than any model choice. The strategies that work.
    8 min
  4. 04
    Building a basic RAG pipeline
    End to end retrieval in the smallest form that actually works, and what to add only when you need it.
    9 min
  5. 05
    When RAG fails
    The specific failure modes of retrieval systems, how to tell them apart, and what fixes each.
    8 min
  6. 06
    Agent memory patterns
    The practical ways agents remember things, from scratch files to vector stores, and when each is right.
    8 min
TRACK 066 Steps

Evaluation and Testing

Knowing whether a change made your agent better, rather than guessing.

  1. 01
    Why you cannot eyeball quality
    Why trying a few prompts feels like testing but is not, and what to do instead.
    7 min
  2. 02
    Building a regression set
    Assemble a small suite of cases that tells you whether a change helped or hurt.
    8 min
  3. 03
    Using a model as a judge
    Grading output with another model, the biases that make it unreliable, and how to correct for them.
    8 min
  4. 04
    Measuring tool use reliability
    Most benchmarks barely test tool calling, yet it is where agents fail most. How to measure it yourself.
    7 min
  5. 05
    Comparing two prompt versions
    How to tell whether version B is genuinely better than version A, rather than luckier.
    7 min
  6. 06
    Tracing and observability
    What to log so that a bad answer in production can actually be diagnosed.
    7 min
TRACK 077 Steps

Production, Cost and Safety

Shipping an agent people depend on without a surprise bill or a security incident.

  1. 01
    Controlling token cost
    Where the money actually goes in an agent workload, and the changes that cut it most.
    8 min
  2. 02
    Latency and streaming
    Why agents feel slow, which parts you can fix, and how streaming changes the experience without changing the speed.
    7 min
  3. 03
    Prompt injection
    Why an agent that reads untrusted content can be turned against you, and what actually mitigates it.
    8 min
  4. 04
    Sandboxing and least privilege
    Containing an agent so that a mistake or an injection is survivable rather than serious.
    8 min
  5. 05
    Handling failures and retries
    Agents fail in ways ordinary software does not. How to build something that degrades sensibly.
    7 min
  6. 06
    Rate limits and scaling
    What actually limits throughput when an agent goes from your laptop to real traffic.
    7 min
  7. 07
    Where to go next
    What to learn once you have a working agent setup, in the order that pays off, and what to safely ignore for now.
    6 min

Test yourself

200 questions across the same seven tracks. Randomised each time, scored, and your attempts are kept so you can compare.