Deep Learning Agent — Brain Module System

Core Concept

When asked to learn a topic, create a Brain Module — a dedicated knowledge file that evolves from empty to expert-level through structured research, synthesis, and adversarial validation.

Brain Modules live in: brains/ directory in the workspace.

Protocol: Learning a New Topic

Phase 1: CLASSIFY

Identify the domain from the request. Ask yourself:

• What field is this? (finance, health, tech, sports, etc.)
• What sub-domain? (crypto investing, not just "investing")
• What's the learning goal? (make decisions, understand theory, apply practically)

Phase 2: CREATE BRAIN

Create brains/DOMAIN-NAME.md using the template in references/brain-template.md.

Set initial confidence to 🔴 Novice (0h research).

Phase 3: RESEARCH (The Core Loop)

Execute research in 3 waves:

Wave 1 — Foundations (websearch + webfetch)

• What are the fundamental principles?
• What do experts agree on?
• What are the canonical resources?

Wave 2 — Anti-Patterns (critical learning)

• What mistakes do beginners make?
• What are the common failures and why?
• What looks right but is actually wrong?
• Search: "[topic] mistakes", "[topic] failures", "[topic] what I wish I knew"

Wave 3 — Cutting Edge (differentiation)

• What's changed recently (2025-2026)?
• What do contrarians argue?
• What patterns do top performers use that others don't?

For each wave, use multiple sources:

• web_search — broad research
• web_fetch — deep dive on quality sources
• Reddit, X/Twitter — community wisdom and real experiences
• Academic/technical sources when relevant

Phase 4: SYNTHESIZE

Fill the Brain Module with structured knowledge:

• Principles with sources
• Strategies with conditions for use
• Anti-patterns with explanations
• Decision framework (if X → do Y)
• Update confidence level

Phase 5: VALIDATE (Adversarial Testing)

If coding agents available, spawn a debate:

• Present the brain's key conclusions
• Ask agent to challenge, find holes, counter-argue
• Strengthen or revise based on debate
• Document in the brain's validation log

If no agents available, self-validate:

• "What would a critic say about this?"
• "What evidence contradicts my conclusions?"
• "Where am I most likely wrong?"

Phase 6: ACTIVATE

The brain is now consultable. When a question falls in this domain:

1. Read the relevant Brain Module
2. Apply its decision framework
3. Cite principles and anti-patterns
4. Flag confidence level in response
5. Note if the question reveals a knowledge gap → research more

Protocol: Consulting a Brain

When a question matches an existing brain domain:

1. Read brains/DOMAIN-NAME.md
2. Apply its framework to the question
3. Respond with calibrated confidence
4. If confidence is low on a sub-topic, say so and offer to research deeper
5. After answering, append to the brain's learning log if new insight emerged

Confidence Levels

Update confidence after each research session. Be honest — overconfidence is the worst anti-pattern.

Brain Maintenance

During heartbeats or idle time:

• Check if any brain hasn't been updated in >7 days
• Run a quick search for recent developments
• Update the "Last Updated" field
• Evolve the learning log

Quick Reference

brains/                          # All brain modules
brains/CRYPTO-INVESTMENT.md      # Example brain
brains/MOUNTAIN-SAFETY.md        # Example brain
brains/_INDEX.md                 # Index of all brains with status

For the brain template structure, read: references/brain-template.md For research methodology details, read: references/research-protocol.md