OSIL Layered Stack

Six layers, each higher layer consumes lower-layer artifacts. L1 is foundational; L6 is observe-only.

graph LR
    subgraph L6["Layer 6 Track-only research-tier"]
        L6A[HyperAgents]
        L6B[AlphaEvolve / SWE-RL]
        L6C[EvoAgentX firehose]
        L6D[GenericAgent / Evolver]
    end

    subgraph L5["Layer 5 Memory Evolution"]
        L5A[Honcho dialectic]
        L5B[Mem0 production]
        L5C[Letta MemGPT]
    end

    subgraph L4["Layer 4 Skill Induction + Autoresearch"]
        L4A[Voyager pattern<br/>auto-skill discovery]
        L4B[Karpathy autoresearch<br/>overnight loops]
        L4C[GenericAgent reference]
    end

    subgraph L3["Layer 3 Prompt Optimization"]
        L3A[DSPy + GEPA]
        L3B[TextGrad]
        L3C[OPRO pattern]
        L3D[PromptBreeder]
        L3E[MIPROv2 fallback]
    end

    subgraph L2["Layer 2 Failure Loops"]
        L2A[Reflexion runner]
        L2B[Self-Refine pattern]
        L2C[CRITIC for code]
    end

    subgraph L1["Layer 1 Capture"]
        L1A[peterskoett self-improvement skill]
        L1B[AGENTS.md formalization]
        L1C[Trajectory capture ShareGPT]
    end

    L1A --> L2A
    L1B --> L3A
    L1C --> L4B
    L1C --> L4A
    L1C --> L3A
    L2A --> L3A
    L3A --> L4A
    L4A --> L5A
    L5A -.feedback.-> L4A
    L6A -.observe.-> L4A

Layer descriptions

  • L1 Capture — passive collection of learnings, AGENTS.md notes, trajectory traces. Foundation; everything depends on this.
  • L2 Failure Loops — generate reflections from failures, store and retrieve. Reduces error rate.
  • L3 Prompt Optimization — GEPA / TextGrad / OPRO / PromptBreeder / MIPROv2. Mutates prompts based on eval data.
  • L4 Skill Induction + Autoresearch — auto-generates new skills from successful traces; runs autoresearch loops on measurable workflows.
  • L5 Memory Evolution — alternative memory backends (Honcho/Mem0/Letta) evaluated against current vec0+FTS5+Voyage-4 hybrid.
  • L6 Track-only — frontier research observed quarterly; not implemented unless validated.