Agent Atlas

Plate I

The Agent Engine

Watch an AI agent turn one sentence into a change to a file. Open any part of the machine to see what it does, then break it on purpose.

Your request

Make the recipe serve instead of 4.

This simulation understands these two requests. A real model accepts any sentence.

THE HARNESSYOUMake the pancakesrecipe serve 6 insteadof 4.THE MODELSealed. Shows what goesin and what comes out.TURN0TOKENS READ0CONTEXT WINDOW55/2000System prompt55PERMISSIONallowdenyTOOLSlist_filesread_fileedit_filePROJECT FOLDERAGENTS.mdrecipes/pancakes.mdrecipes/flatbread.mdANSWERWaiting for the loop to end.THE HARNESSYOUMake the pancakesrecipe serve 6instead of 4.THE MODELSealed. Shows what goesin and what comes out.TURN0TOKENS READ0CONTEXT WINDOW55/2000System prompt55PERMISSIONallowdenyTOOLSlist_filesread_fileedit_filePROJECT FOLDERAGENTS.mdrecipes/pancakes.mdrecipes/flatbread.mdANSWERWaiting for the loop to end.
Step 1 of 17

Load The harness writes the system prompt into the context.

Keys: ← → step, P play or pause, Home start. Select any part of the drawing to look inside it.

Select a part of the drawing, or pick one here:

Experiments

Change one thing and the machine runs again.

The ledger

The same run as text. It is complete without the animation, and each step can be opened.

  1. The harness writes the system prompt into the context.
  2. The harness describes the three tools the model may ask for. These descriptions take up context too.
  3. The harness finds AGENTS.md in the project and loads it before the request, so its rules apply from the start.
  4. Your request joins the context. The model has not seen anything yet.
  5. Turn 1. The model reads the whole context and replies with a tool request, not prose.
    Data
    {
      "tool": "list_files",
      "input": {}
    }
  6. list_files only reads, so the harness lets it run without asking you.
  7. The tool runs and its output is added to the context.
    Data
    AGENTS.md
    recipes/pancakes.md
    recipes/flatbread.md
  8. Turn 2. The model asks to read recipes/pancakes.md.
    Data
    {
      "tool": "read_file",
      "input": {
        "path": "recipes/pancakes.md"
      }
    }
  9. read_file only reads, so the harness lets it run without asking you.
  10. The tool runs and its output is added to the context.
    Data
    # Saturday pancakes
    
    My grandmother made these every Saturday in a pan older than my father.
    She never measured anything, so this recipe is my best guess from watching her.
    The trick is to stop stirring while the batter still looks a little lumpy.
    Let it rest while the pan heats. Nobody knows why that works, but it does.
    Flip once, when bubbles stay open on top. Never press them flat.
    Serve with butter and whatever fruit is about to go off.
    
    Serves 4. Takes 20 minutes.
    
    ## Ingredients
    - 2 cups flour (250 g)
    - 1 ½ cups milk (360 ml)
    - 2 eggs
    - 2 tbsp sugar (25 g)
    - 2 tsp baking powder (10 g)
    - ½ tsp salt (3 g)
    - 3 tbsp melted butter (45 g)
    
    ## Method
    1. Whisk the dry ingredients.
    2. Beat the eggs into the milk and butter.
    3. Fold wet into dry. Stop while it is still lumpy.
    4. Rest 5 minutes, then cook on a medium pan.
  11. Turn 3. The model asks to replace the serving line and amounts in recipes/pancakes.md.
    Data
    {
      "tool": "edit_file",
      "input": {
        "path": "recipes/pancakes.md",
        "old_text": "Serves 4. Takes 20 minutes.\n\n## Ingredients\n- 2 cups flour (250 g)\n- 1 ½ cups milk (360 ml)\n- 2 eggs\n- 2 tbsp sugar (25 g)\n- 2 tsp baking powder (10 g)\n- ½ tsp salt (3 g)\n- 3 tbsp melted butter (45 g)",
        "new_text": "Serves 6. Takes 20 minutes.\n\n## Ingredients\n- 375 g flour\n- 540 ml milk\n- 3 eggs\n- 40 g sugar\n- 15 g baking powder\n- 5 g salt\n- 70 g melted butter"
      }
    }
  12. Editing changes your files, so the harness stops and asks you. You allow it.
  13. The harness writes recipes/pancakes.md and returns the outcome to the context.
    Data
    Edited recipes/pancakes.md.
  14. Turn 4. Before answering, the model reads back the lines it changed.
    Data
    {
      "tool": "read_file",
      "input": {
        "path": "recipes/pancakes.md",
        "offset": 9,
        "limit": 10
      }
    }
  15. read_file only reads, so the harness lets it run without asking you.
  16. The tool runs and its output is added to the context.
    Data
    Serves 6. Takes 20 minutes.
    
    ## Ingredients
    - 375 g flour
    - 540 ml milk
    - 3 eggs
    - 40 g sugar
    - 15 g baking powder
    - 5 g salt
    - 70 g melted butter
  17. The lines it read back match the lines it asked to write, so the model replies in words and the loop ends.
    Data
    Done. recipes/pancakes.md now serves 6. Amounts are in grams and ml, rounded as AGENTS.md asks.

Outcome: Done. recipes/pancakes.md now serves 6. Amounts are in grams and ml, rounded as AGENTS.md asks.

Where to go next

  • II. The Scaffold. What an agent reads before it reads your request: AGENTS.md, CLAUDE.md, skills, hooks and memory, taken apart.
  • III. The Command Deck. A practice terminal. Move around a project the way an agent does.
  • IV. A Life in Context. What happens to an agent from a fresh start to/clear, with a real session and the habits that keep it sharp.
  • V. The Sampler. How a model's probabilities become one word, with temperature and top-p on real numbers from a small model.
  • VI. The Machine Room. Follow one request from a laptop to a lab's data center and back, then run it on a model at home.
  • The Builders. How the people building with agents work: their machines, their budgets, their habits, and where you can start.
  • The Frontier. What eight labs publish about their leading models, dated, with every gap marked.
  • The Atlas. Every idea on these plates, mapped and linked, with a twelve-step path for beginners.
  • Field guide to agent files. What each file is for, who reads it, and when.