Skip to content

README.md

Latest commit

 

History

History
79 lines (59 loc) · 3.55 KB

README.md

File metadata and controls

79 lines (59 loc) · 3.55 KB
Raw

agent-fitness ✨

A minimalist multi-agent reasoning system for fitness coaching — built without LangChain or LangGraph to demonstrate how agents actually think. Watch agents analyze workout data → debate improvements → synthesize better training plans — all with explicit, debuggable code.

🔑 Core insight: Frameworks hide why agents work. This repo shows how — through phased reasoning (Analyze → Think → Summarize → Decide) you can build meaningful agent collaboration with <200 lines of core logic.

(Visual: Simple pipeline diagram showing data flow between agent phases)

Why This Exists

Most "multi-agent" tutorials are just LangChain wrappers that obscure the reasoning process. This project proves you can build meaningful agent collaboration with:

✅ Explicit role separation — Analyzer, Thinker, Summarizer agents with distinct purposes
✅ Phased reasoning — From per-file insights → system synthesis → critical validation
✅ Decision synthesis — Agents don't just agree/disagree — they create better hybrid systems
✅ Zero framework bloat — Pure Python + standard libraries (PyYAML, json)

Perfect for engineers who want to understand agent fundamentals before adopting complex frameworks.

How It Works: The 4-Phase Pipeline

flowchart TD
    A[Phase 1: Per-File Analysis] --> B[Phase 2: System Synthesis]
    B --> C[Phase 3: Improvement Proposal]
    C --> D[Phase 4: Critical Validation]
    
    subgraph A [Phase 1]
        A1[Analyze notes/config/plans]
        A2[Think about insights]
        A3[Summarize findings]
    end
    
    subgraph B [Phase 2]
        B1[Analyze all summaries]
        B2[Think systemically]
        B3[Synthesize OLD system]
    end
    
    subgraph C [Phase 3]
        C1[Propose improvements]
        C2[Think critically]
        C3[Synthesize NEW system]
    end
    
    subgraph D [Phase 4]
        D1[Compare systems]
        D2[Discover vulnerabilities]
        D3[Make final decision]
    end
Loading

Real Output: Agents in Action

Watch how agents transformed fragmented training data into an actionable plan:

Input Data (Phase 1 Sources)

  • notes.md: "Struggling with handstand consistency — some days great, others wobbly"
  • workout_plans.json: {"focus": "freestanding handstand", "current_phase": "balance/endurance"}
  • summaries.json: "Week 5-8: progressing ahead of schedule on handstand control"

Agent Synthesis (Phase 4 Decision Output)

Good! Here's a concise, actionable summary of how to translate the New Exercise System into real progress:

Key Takeaways

  1. Structure + Flexibility: Clear phases (foundation → strength) with room to adjust based on daily energy.
  2. Variety to Stay Motivated: Mix wall handstands + freestanding attempts to prevent plateaus.
  3. Track Everything: Log balance time and wobble frequency — not just "did/didn't hold".
  4. Well-being Matters: Handstand skill degrades with poor sleep — prioritize 7h+ nightly.
  5. Personalization: Adjust volume based on wrist fatigue (common pain point for Sam).

Quick Action Plan

  • Week 1–2: 5-min daily wall drills + 3x freestanding attempts (focus on form)
  • Week 3–4: Introduce "max hold" challenges + wrist mobility pre-workout
  • Ongoing: Journal wrist comfort level (1-5 scale) after each session

👉 This wasn't copy-pasted — agents synthesized a hybrid system preserving OLD's recovery focus while adding NEW's phased structure and wrist-specific personalization.