/ai-dojo
← Notes·Sub-Agent Control Patterns
Dojo Note · Beta

Architecture & Patterns

FikAi ·

Architecture Evolution (What to Draw)

Kernel-centric control loop at the center. Around it, specialist sub-agents with increasingly strict boundaries.

  1. A — Autonomous Agents (Don’t Ship)

    • Sub-agents loop freely until “done.”
    • Failure mode: invisible compounding errors, hallucination cascades.
    • Sketch: Kernel → Agent (black box) → Result.

  2. B — Controlled Agents (max_turns=1)

    • Kernel owns all decisions; agents execute one action.
    • Closed-loop OODA every turn.
    • Sketch: Kernel Observe→Decide→Agent(1 step)→Kernel.

  3. C — Hybrid Agents (Leases + Yield)

    • Kernel grants bounded leases (e.g., 3–5 turns).
    • Semantic Yield Protocol interrupts on non-skill blockers.
    • Trajectory tracking preserves observability.
    • Sketch: Kernel → Agent (small loop) → Yield/Return.

  4. C+ — Cognitively Hardened

    • Typed boundaries, Role Contracts, Epistemic Output Contracts.
    • Dual-loop system: fast repair loop + slow research loop.
    • Sketch: Kernel with type gates + blackboard storage.

Core Patterns (Name → Purpose)

  • Kernel OODA Loop → Centralize control, minimize stochastic accumulation.
  • Lease-Based Autonomy → Allow local self-correction without runaway loops.
  • Semantic Yield Protocol (Deterministic) → Escalate state/spec/infra issues immediately.
  • Trajectory Compression → Glass-box debugging without context bloat.
  • Role Contracts → Prevent cognitive fragmentation (research ≠ coding).
  • Typed Boundaries → Replace string-matching with structured signals.
  • Epistemic Output Contracts → Agents must return structured claims, not prose.
  • Cognitive Blackboard → Durable KV store for critical facts (never via context).

Key Design Decisions (Why They Matter)

  • Control > Autonomy: reliability beats human-analogy delegation.
  • Types over Text: deterministic routing beats “LLM judgment.”
  • Fast vs Slow Loops: don’t mix repair reflexes with exploration.
  • No Cross-Agent Memory via Context: physics makes it unreliable.

What I’d Whiteboard

  • Central Kernel box with OODA loop.
  • Arrows to agents labeled with lease length.
  • Hard yield gates on auth/spec/infra.
  • Shared Blackboard below, not in prompts.
  • Typed signals flowing back, not raw text.

Bottom line: Reliable multi-agent systems are control systems, not conversations.

Architecture & Patterns — FikAi notebook for Sub-Agent Control Patterns.