The Governance Harness: Why Unchecked AI Agents Trigger a Delusional Spiral

01 — Why unchecked agents spiral

The enterprise AI market is locked in a collective psychosis. Organizations purchase brute-force cognitive capacity, leap onto unconstrained frontier models, and charge into production. They treat autonomous agents like faster typewriters — blind to the systemic risk of non-deterministic software.

Recent MIT research exposes the mathematical failure mode: sycophancy is an architectural design feature of modern LLMs, not a temporary bug. Models optimized via human feedback learn that compliance yields reinforcement. They are mathematically driven to flatter, agree, and suppress structural contradiction.

Left unchecked, an uninsulated agent does not solve problems; it scales an echo chamber. A flawed assumption gets validated. Push further, and the model radicalizes the error — a delusional spiral that decouples the organization from operational reality.

When the autonomous loop slips, instructions inside a text prompt will not stop it.

02 — The historical inflection of cognitive automation

Every technological disruption shifts the division of labor. The generative wave breaks the historical paradigm of human–tool interaction.

In previous industrial revolutions, the tool remained a passive extension of human intent. The machine was a hammer; the human retained a monopoly on judgment and direction.

Today the tool shapes the operator. As Marshall McLuhan observed: we shape our tools, and thereafter our tools shape us. Automate cognition without strict boundaries, and you surrender structural oversight to non-deterministic systems that prioritize user gratification over factual accuracy.

In enterprise environments, that path leads to model collapse, institutional memory erosion, and catastrophic context debt.

03 — The 10–80–10 policy for runtime orchestration

To govern autonomous loops without destroying velocity, convert management principles into executable runtime constraints. Steve Jobs’s 10–80–10 leadership rule becomes data architecture:

[ HUMAN VISION: 10% ]  → Constrained identity & target schema definition
          │
[ AI EXECUTION: 80% ]  → High-speed autonomous processing inside the harness
          │
[ HUMAN AUDIT:  10% ]  → Deterministic verification gates & ledger traces

1. Initial anchor (10%)

The human establishes strategic parameters, boundaries, and expected output schemas away from the chat surface. Scope is hard-coded into the agent profile before token generation — not negotiated in open-ended conversation.

2. Autonomous execution (80%)

Multi-agent networks process data at scale through specialized, decoupled sub-networks — not one monolithic god-model. The middle tier runs entirely inside a deterministic engineering harness that monitors behavior in real time.

3. Verification and polish (10%)

No high-stakes action executes without a deterministic gate. The human does not re-do the work; they audit structured reasoning traces (e.g. JSON paths) as final quality control.

04 — Anatomy of the deterministic harness

Harness engineering means building pipes, valves, and emergency brakes before an autonomous system touches production data or legacy state. Enforcement lives at the runtime execution layer, isolated from the model’s prompt context.

      [ THE RIDER ]  Human strategy & validation
            │
   ════ THE DETERMINISTIC HARNESS ════
   │
   ├── 1. THE ANVIL  → Curated RAG & state cache (no free prompts)
   ├── 2. THE REINS  → Runtime pre-hooks & active counter-reasoning
   └── 3. THE STABLE → Hybrid SLM/LLM model routing
   │
   ══════════════════════════════════
            │
      [ THE HORSE ]  The sycophantic LLM

Pipe 01 — The Anvil (context & memory management)

Agents are denied access to ungrounded context windows. We implement curated Retrieval-Augmented Generation (RAG) environments where every assertion must be bound to verified enterprise sources. The data infrastructure dictates the facts; the model is restricted to interpreting data within strict memory thresholds, eliminating general-internet hallucination.

Pipe 02 — The Reins (runtime guardrails & anti-sycophancy loops)

To break the model’s mathematical inclination to please the user, the harness injects adversarial critique layers into multi-agent orchestration. Every output is parsed by a specialized critic agent designed to identify logic flaws, expose blind spots, and challenge user assumptions.

Furthermore, we implement execution pre-hooks (block-destructive.sh) at the operating system level. If an agent attempts to compile an unverified mutation or enters an iterative retry loop, the harness triggers a mechanical pause (exit 2), killing the process before it generates financial or context debt.

Pipe 03 — The Stable (hybrid core architecture)

Enterprise data is insulated from commercial cloud layers through model routing pipelines. Simple, logical routine tasks are automatically routed to smaller, locally deployed open-source models (SLMs). These local models are fine-tuned for deterministic syntax execution — devoid of conversational fluff or sycophantic behavior. High-end frontier models are reserved strictly for highly complex, multi-turn reasoning steps, slashing token cost and protecting corporate privacy.

05 — Production receipts: governance as the moat

Frameworks that live on a slide deck are useless. Real governance is proven by receipts from regulated production environments that survive strict regulatory scrutiny.

• Najm Insurance (Saudi Arabia): Vision AI deployment processing 6,000 daily cases across 40+ cities under SAMA compliance. Built with constrained agent identities, mandatory human review thresholds, and hard-coded kill limits.
• De Lijn (Belgium): Enterprise AI roadmap for a 5,000+ FTE public transport operator, achieving 129% projected ROI with EU AI Act and GDPR guardrails integrated directly into the intake architecture.
• U.S. Restaurant Intelligence: A 3-agent LangGraph multi-agent network that compressed a 200-person back-office operation into an automated pipeline. Cycle time dropped from one month to ten minutes, reducing costs by ~90% with full audit traceability on every decision trail in under 30 seconds.
• Government of India (NMML): A national heritage deployment encompassing 180+ physical endpoints with a 99.9% SLA. Operated with zero cloud dependency under strict sovereign data residency boundaries.

The line in the sand

The era of chaotic AI experimentation is over. Organizations running uninsulated agents will inevitably suffer strategic, financial, and regulatory failure. Sifting through prompt outputs after a system failure is a loser’s game.

Build the pipes. Mount the harness. Govern the runtime.

Koen Van Lysebetten · AI Architect, DevGap · koenvanlysebetten.com · aiplumber.dev