Quick Reference

TL;DR (30 seconds)

Autonomous sub-agents fail by physics, not prompts. Centralize control, bound autonomy, and harden agent boundaries with typed signals and role contracts to reach ~99% reliability. Treat multi-agent systems as control systems with error injection at every boundary.

Architecture Spectrum (know this table)

• A – Autonomous loops: Full freedom → 38% failure. Never ship.
• B – Controlled (max_turns=1): Kernel decides every step → 91% success.
• C – Hybrid (lease + yield): Limited autonomy within competency → 93–95%.
• C+ – Cognitively hardened: Typed boundaries, dual-loop, epistemic contracts → 99.2%.
  → See Fig. 1 (Architecture evolution).

Non‑Negotiable Laws (from Vol. 1, applied here)

• Finite Attention: Cross-agent recall multiplies losses.
  • Formula: P_cross = P_A × P_serialization × P_B → example 0.38 (Sec. “Law 1 in Multi-Agent”).
• Stochastic Accumulation: Errors compound across agents.
  • P_fail = 1 − Π(1 − p_i) → 4 agents @5% ⇒ 18.5% (Law 2 section).
• Entropic Expansion: Kernel context overflows without compression.
  • Use ≥90% compression on trajectories (Law 3).

Control Patterns to Use

• Kernel OODA loop: Observe → Budget → Orient → Decide (1 action) → Checkpoint (Sec. Arch B).
• max_turns=1 default. Any loop >1 must be leased and yieldable.
• Hybrid lease: N ≤ 5 turns max; cannot self-extend (Arch C).

Yield Early, Deterministically

• Semantic Yield Protocol (hardcoded, not LLM):
  • Yield immediately on 401/403, timeouts, connection refused, 405, missing creds.
  • See Python snippet in Arch C.

Role Discipline (Cognitive Cohesion)

• Role Contracts: Epistemic domain + tools + yield conditions (Sec. “Role Contract Theory”).
• Never mix domains in one agent. Quality drops ~30% for 2 domains (Q ∝ D^−α).
• Enforce by tool absence, not instructions.

Typed Boundaries (don’t skip)

• No string-matching for failures.
• Use typed results, typed failures, epistemic output contracts.
• Prevent partial success (clean exit ≠ mission success).

Where to Look

• Why autonomy fails → Architecture A
• Why max_turns=1 works → Architecture B
• Leases + yield → Architecture C
• Cognitive hardening patterns → Sections 7–12

Bottom line: If an agent can loop, it must be leased, observable, typed, and yieldable—or it will fail at scale.

---

Quick Reference — FikAi notebook for Sub-Agent Control Patterns.