Protocol 03 of 18Track, MindsetHow we think

Agents, the folder is the agent

An agent is not a magical entity. It is a folder: a role, skills, context, routines, guardrails, and a Definition of Done. Structure the folder well and the agent works. Structure it badly and it does not.

← All 18 protocols
Protocol 02The stack, Claude plus GitHub plus Vercel plus SupabaseAll 18Protocol 04The build team, humans plus AI
Why this matters

The pain it
solves

Agent is the most overused word in AI right now. Most people imagine a black-box autonomous robot. That misunderstanding is expensive. It makes people scared of agents (too magical), disappointed by them (not magical enough), or sold things that are not really agents.

Once you see that an agent is a folder, the fear and mystique go away. You can read it. Edit it. Copy it. Hand one to a teammate. Build a new one in fifteen minutes.

The teaching

What this
actually is

What is actually in the folder

An agent is not a magical entity. It is a folder on your computer that contains six kinds of file. The LLM is the same Claude everyone else uses. The folder is what makes Claude behave like a specialist.

  • Role definition
    A markdown file (often called AGENT.md or named after the agent) that tells the LLM who it is, what it does, what to avoid, and what tone to use. The shortest version of a job description.
  • Skills
    Reusable prompts the agent invokes for specific tasks. Slash commands. Short, named playbooks. The verbs of the agent.
  • Context
    The files the agent reads before it starts. The product plan for the developer. The QA plan for QA. The design system for anyone who writes UI. Anything the agent should know before the first line of work.
  • Routines
    What the agent does on a schedule or a trigger, without being asked. Every morning: read what shipped and propose the day. Every commit: run the tests. Every new lead: score it. A routine is a skill plus a trigger.
  • Guardrails
    What it may never do without you. Never email a customer. Never delete a row. Never push to main. Written down in the folder, so the rule survives the session and the model.
  • Definition of Done
    How the agent knows it is finished. A checklist it can test, with evidence. Without it, an agent either stops too early or never stops. With it, it stops, reports, and waits.

Why the folder framing matters

Once you see that an agent is a folder, the fear and the mystique both go away. You can read it. Edit it. Copy it. Hand one to a teammate. Make a new one in fifteen minutes by copying an existing one and changing the role.

The opposite framing, the magic-black-box framing, is what makes people scared of agents, disappointed in them, or sold things that are not really agents at all. If you cannot open the folder, you do not own the agent.

How to configure one

Start from an agent that exists. Copy its folder. Rewrite the role in three sentences. List the skills it needs and delete the ones it does not. Point its context at the files for this job. Give it one routine, one guardrail, and a Definition of Done it can test. Run it on a small prompt. Read the output. Edit the folder. Run it again.

Every agent you will ever build is that loop. The folder is the agent; the loop is how you raise it.

Try it yourself20 minutes

Read, edit, and rerun one agent in 20 minutes

If you have the IL agent team installed (from the retreat or the template repo), use the developer agent. If not, copy any agent's folder from the template into a new project to follow along.

  1. Step 01
    Open the agent's folder

    In your terminal, cd into .claude/agents/developer (or wherever your agent lives). Run ls. You will see a small set of markdown files. Open them in your editor.

  2. Step 02
    Read the role definition

    Open the file that defines the agent (typically AGENT.md, developer.md, or a file with frontmatter). Notice how short and specific it is. Find the six parts: role, skills, context, routines, guardrails, Definition of Done. If one is missing, that is your first edit.

  3. Step 03
    Run the agent on a small prompt

    In Claude Code, ask the developer agent: "Add a 50-word intro paragraph to the homepage and stop when it renders." Save the output.

  4. Step 04
    Edit the role definition

    Change one line in the role file: a guardrail ("never edit files outside src/") or the Definition of Done ("finished means the change is live on the Vercel preview"). Save.

  5. Step 05
    Run the same prompt again

    Ask the same agent the same prompt. The output will be visibly different. Same Claude, different folder, different result.

Outcome

You have seen, with your own eyes, that the agent is the folder. Same model, different folder, different specialist. The black-box mystique is gone for good.

Official resources

Straight from
the source

Anthropicdoc
Claude Code subagents
agents.mddoc
AGENTS.md, an open standard
Anthropicdoc
Claude Code slash commands
What you walk out with

By the end of this
protocol

At the retreat

You learn it by
doing it

You open an agent, change one line of its role definition, run it twice on the same prompt, and watch the same model produce two different outputs because the folder told it to.

Bring it to your companySee all 18 protocols
Connects to

Other protocols this
compounds with

Protocol 04
The build team
Protocol 08
ABC, always be cataloguing
Protocol 15
Agents in the cloud
← Previous, Protocol 02
The stack, Claude plus GitHub plus Vercel plus Supabase
Next, Protocol 04 →
The build team, humans plus AI