Agentic Coding: Run Coding Agents Like an Operator
How to actually drive coding agents at work: pick a harness, control the context window, write rules and skills that stick, wire hooks and MCP, and run whole tickets through parallel multi-agent workflows.
You ship code and use an agent casually. This course is the machinery under it: what the model sees, what you can put in front of it, and how to run many agents at once without losing control of the result.
It is the onboarding I give engineers joining my team, written down.
What you'll be able to do
- Pick a harness and a subscription on purpose
- Read your own session: every tool definition, rule file and skill that loaded
- Cut your context window to what the model needs, and spot when it goes bad
- Keep one set of agent rules every harness on your machine reads
- Write skills the agent loads only when it should
- Move format, lint and checks off the model into hooks
- Route every external service through one tool layer, not a dozen MCP servers
- Run a real ticket end to end: research, plan, parallel implementation, review, browser QA, pull request
What this course assumes
You write code for a living and have opened Claude Code or Codex once. Examples use Claude Code, and say where a feature is Anthropic-only.
How it works
Every lesson is free and open, no account needed. Read them in order; each builds on the one before, and your progress saves in your browser.
Lessons
- 01
Choose Your Coding Agent
3 minPick a harness by working backwards from the subscription that is allowed to drive it, and know which features you lose when you switch.
- 02
How the Model Actually Works
3 minThe model is stateless and emits only text: see the messages array your harness resends every turn, and how to budget a plan with a weekly cap.
- 03
Rules Files and Scope
4 minWrite your agent rules once in ~/.agents/AGENTS.md, symlink every harness to it, and know exactly which files each harness loads and in what order.
- 04
Own Your Context Window
5 minRead what your agent actually loads, cut the tool definitions and bundled extras you never use, and start fresh before quality drops.
- 05
Skills: Teach the Agent Your Way of Working
3 minPackage your repeated instructions as skills: the SKILL.md format, progressive disclosure, prompt vs router vs script-backed shapes, and how to install them safely.
- 06
Hooks: Move Work Off the Model
3 minWire format, lint and guardrail commands into agent lifecycle events so the machine does them every turn instead of the model remembering to.
- 07
MCP Without Tool Bloat
5 minConnect Jira, Notion, PostHog, Stripe and the rest through one gateway, so hundreds of endpoints cost a handful of tool definitions.
- 08
Voice Dictation
3 minDictate prompts instead of typing them: a local speech-to-text app for free, Groq for cents, and one rules line that makes the agent forgive homophones.
- 09
Dynamic Workflows
3 minType ultracode and the model writes a script that spawns sub-agents: phases in sequence, agents in parallel, parent context window almost empty.
- 10
Workflows in Action
4 minOne real ticket end to end: a research prompt that produces a reviewable plan, an execution prompt that ships it, and the diff review that catches what the summary hides.