70. Nested Learning

Mini-Project: Nested Learning: 3 Loops Where Each Level's Output Changes the Next Iteration's Behavior

A three-level learning system where an inner loop retries low-scoring answers with chain-of-thought, a middle loop permanently adds CoT to the prompt when it consistently helps, and an outer loop adds a calculator tool when performance plateaus across epochs.

View on GitHub

---

Description

Nested Learning creates a hierarchy of learning loops operating at different time scales. The inner loop (fast) optimizes within a single task (e.g., prompt refinement during execution). The middle loop (medium) optimizes across tasks (e.g., updating strategies based on daily performance). The outer loop (slow) optimizes the agent architecture itself (e.g., monthly review of tool selection and workflow design).

How It Works

flowchart TD
    A[Outer Loop: Architecture Evolution - Monthly]
    A --> B[Middle Loop: Strategy Optimization - Daily]
    B --> C[Inner Loop: In-Task Adaptation - Per Task]
    C --> D[Execute Task]
    D --> E[Task Feedback]
    E --> C
    E --> B
    E --> A

    style A fill:#F44336,color:#fff
    style B fill:#FF9800,color:#fff
    style C fill:#2196F3,color:#fff