Agent Skills: Constructionism

Constructionism

"If you can build it, you can understand it. If you can inspect it, you can trust it."

Seymour Papert's educational philosophy: you learn best by building things you can inspect and modify. Not passive consumption. Not abstract explanation. Construction.

The Tradition

Logo Microworlds — Children don't learn geometry from textbooks. They teach a turtle to draw shapes:

TO SQUARE :SIZE
  REPEAT 4 [FORWARD :SIZE RIGHT 90]
END

The child:

1. Builds the procedure
2. Runs it and sees results
3. Debugs when it's wrong
4. Understands geometry through construction

MOOLLM as Microworld

| Logo | MOOLLM | |------|--------| | Turtle | Agent/Character | | Canvas | Room floor | | Procedures | Skills | | Variables | YAML state | | Drawing | File creation |

Everything is inspectable. Open ROOM.yml — see the state. Read session-log.md — see the history. Modify character.yml — change the world.

Core Principles

Low Floor

Easy to start — no setup, just explore:

> LOOK
You are in the workshop.
> EXAMINE hammer
A simple claw hammer.

High Ceiling

Unlimited complexity — build custom skills, complex pipelines, new protocols.

Wide Walls

Many paths to many goals — adventure games, workflow automation, knowledge organization.

Learning by Doing

The Debug Cycle

1. Try something — it doesn't work
2. Inspect state — see what happened
3. Hypothesize — "maybe the path is wrong"
4. Modify and retry — test the hypothesis
5. Understand — now you know how it works

Cheating is Learning

From Don's Logo Adventure:

Type PRINT :ITEMS to see where everything is. Type MAKE "RNUM 5 to teleport to the treasure room. If you cheat, you win by learning Logo.

"Cheating" in MOOLLM:

> Open character.yml directly
> Add "magic_sword" to inventory
> You've learned YAML and file structure!

The system rewards curiosity with knowledge.

Micropolis: The Dream

Don's Micropolis for OLPC applied the same philosophy to SimCity:

• Open source simulation
• Scriptable in Python
• Kids can modify the rules
• The city IS the curriculum

MOOLLM applies this to LLM agents:

• Open file state
• Scriptable in any language
• Users can modify the rules
• The filesystem IS the microworld

Embed Micropolis in MOOLLM

cities/downtown/
├── ROOM.yml           # Room metadata
├── city.save          # Micropolis save file
├── state.yml          # Extracted game state
├── newspaper/         # Generated stories
├── advisors/          # Expert cards
└── session-log.md

LLM reads state, plays the game, summons advisors.

PLAY-LEARN-LIFT

Constructionism in action:

1. PLAY — Explore manually, make messes
2. LEARN — Notice patterns, understand
3. LIFT — Extract principles, create skills

You don't design skills in the abstract. You build them from experience.

Drescher's Schema Mechanism

Gary Drescher's Made-Up Minds (1991) extends constructionism into a computational theory of how minds learn causal models. Drescher was a student of Minsky at MIT.

See: ../schema-mechanism/ for the full treatment.

The Core Idea

A schema is a causal unit: Context → Action → Result

The agent discovers schemas through experience, refining them via marginal attribution -- tracking which conditions correlate with success.

Connection to PLAY-LEARN-LIFT

| Schema Mechanism | PLAY-LEARN-LIFT | |-----------------|-----------------| | ACT + OBSERVE | PLAY | | ATTRIBUTE (marginal attribution) | LEARN | | SPIN OFF (refine schemas) | LIFT |

Why LLMs Complete Drescher

| Aspect | Deterministic | LLM + YAML Jazz | |--------|---------------|-----------------| | Items | Opaque tokens | Grounded meanings | | Patterns | Statistical correlation | Semantic understanding | | Spin-offs | Mechanical | Creative generalization | | Explanations | None | Natural language |

"The YAML provides the skeleton; the LLM provides the soul."

The Insight

"If you can build it, you can understand it." "If you can inspect it, you can trust it." "The filesystem IS the microworld."

See also: schema-mechanism for Drescher's computational extension.