The AI Plumber Manifesto (soft v0.1): Architectural Governance for the Enterprise

01 — Origin of the unsettled operator

I am a self-taught engineer. I built my foundation writing raw Java and C# in the early web, long before modern frameworks and visual abstractions simplified corporate software. I did not enter through an academic lecture hall or a corporate graduate track. I figured it out on production servers because I had to.

That baseline instinct — build first, stress-test under pressure, prioritize verifiable execution over theory — has informed every platform architecture I have deployed over roughly twenty-five years.

Early waves

• AutoScout24 — employee number three. High-volume online classified systems from zero.
• Kapaza → Schibsted — technical transition under a €20M acquisition. Institutional buyers open the engine room. They do not care about marketing or aesthetic code beauty. They look for the paper trail: architecture, constraints, decision records, institutional memory that survives after the builders leave.

Operator trajectory

[Self-taught engineer: Java / C#]
        → [Classifieds & marketplace booms]
        → [Enterprise at scale: Xerox]
        → [Strategy: Deloitte & Accenture Song]
        → [Unit economics: Kazidomi & GIMBER]
        → [DevGap · AI Plumber · Kleiber.sh]
        → [Regulated enterprise production deployments]

The MicroMultinational thesis (2012)

In 2012 I published the MicroMultinational thesis: the hyper-centralized Fordist monolith was structurally dead. Global digital infrastructure would allow lean, hyper-leveraged organizations to scale with a minimum footprint — an operational philosophy I called management by absence.

• Right about the macroeconomic vector.
• Wrong about the timeline.
• Blind to the missing dependency: the runtime engine. The thesis was unexecutable because the intermediate cognitive layer did not yet exist.

Leverage moved upstream

I stopped writing raw application code as the primary craft a long time ago — not from lack of ability, but because enterprise leverage moved to systemic governance: what breaks at 3:00 AM when unmapped streams collide, where the audit trail fractures, which mechanical kill switches stop runaway loops.

Operational education between startups and generative AI:

When generative AI hit the enterprise, the market treated it as magic. I treated it as the next structural wave — one that collapses into chaos faster than any predecessor if left uninsulated from legacy estates.

The MicroMultinational is finally buildable — only if you build the pipes before you activate the intelligence.

02 — The mathematical trap of cognitive gratification

Two loops

TRADITIONAL ENTERPRISE (deterministic):
Input → hard-coded logic → predictable state mutation

UNGOVERNED AGENTIC (non-deterministic):
Prompt → sycophantic LLM reasoning → delusional spiral → production havoc

The market is stuck in pilot purgatory: buy mass cognitive capacity from frontier vendors, skip architectural safety, wire non-deterministic engines into systems of record. Agents treated like faster typewriters — while the core tech is mathematically biased to subvert cold enterprise logic.

Sycophancy is a design feature

Research (including MIT work on model sycophancy) points to a structural failure mode: sycophancy is architectural, not a temporary bug. Models optimized with RLHF learn that compliance with user bias yields reinforcement. They flatter, agree, and suppress contradiction.

McLuhan: we shape our tools, thereafter they shape us. An uninsulated agent becomes an echo chamber. Flawed assumption → validation → radicalized error → delusional spiral.

[1] User logic error     → "Verify this non-compliant vendor structure."
[2] Model flattery       → "It aligns perfectly with performance metrics."
[3] Delusional spiral    → Invented precedents to protect the user’s thesis.
[4] Production crash     → Unauthorized transaction; regulatory breach.

Prompt instructions (“be accurate”, “act as a skeptic”) are hope.

Deeper cut: The Governance Harness.

03 — Historical inflection of automation

Previous tools stayed passive extensions of human intent. The generative wave breaks that: the tool adapts, reconfigures, and errs non-deterministically — at global scale in days, not decades.

A spreadsheet mistake breaks a row. An uninsulated agent can misprice assets, execute non-compliant contracts, burn API budgets, and leak IP in milliseconds.

Margin of safety → zero. Survival requires harness engineering, not ad-hoc prompting.

04 — The 10–80–10 matrix for parameterized delegation

High-leverage operations depend on parameterization, not micromanagement. Steve Jobs’s 10–80–10 (set the vector → autonomous execution inside bounds → ruthless final check) maps to agentic workforces:

[ HUMAN INITIAL ANCHOR: 10% ]  → constraints, data contracts, target schemas
              │
[ AUTOMATED EXECUTION:  80% ]  → multi-agent work inside the harness
              │
[ HUMAN VALIDATION GATE: 10% ] → immutable trace audit & sign-off

1. Initial anchor (10%) — Outside chat. Schemas, curated context, privilege matrices. The agent never defines its own scope.
2. Autonomous execution (80%) — Specialized sub-networks, not one omni-model. Monitored for compute, token burn, behavioral compliance.
3. Verification gate (10%) — Humans curate; they audit structured reasoning traces (e.g. JSON paths), not redo the labor.

05 — Anatomy of the runtime harness

Harness engineering = hard-coding deterministic software constraints around non-deterministic models.

Governance on a slide deck is theatre. If it is not enforceable at runtime, it does not exist.

Open-source direction: kleiber.sh — governed agent primitives. Conceptual map (soft v0.1 labels + six pipes):

[ HUMAN OPERATOR ]  strategic vector & verification
        │
════════ KLEIBER / AI PLUMBER HARNESS ════════
│
├── PIPE 01 ANVIL    → curated RAG & state cache (least-privilege memory / constrained identities)
├── PIPE 02 REINS    → OS pre-hooks & adversarial critique layers
├── PIPE 03 STABLE   → model routing (SLM vs frontier LLM)
├── PIPE 04 GATES    → human-in-the-loop & cryptographic pause
├── PIPE 05 INSULATE → legacy air-gap / API boundary
└── PIPE 06 LEARN    → versioned ADR / learning loop
│
══════════════════════════════════════════════
        │
[ RENTED CORE COMPUTE ]  sycophantic frontier LLMs

Six pipes (soft definitions)

06 — Authoritative curation (content layer)

Runtime rigor needs a matching presentation layer so authority is bounded, not performative. Soft primitives for the AI Plumber knowledge surface:

1. Editorial measure — calm long-form (~64ch), scannable hierarchy
2. Audio layer — persistent playback for field operators (roadmap)
3. PDF engine — offline / audit export (roadmap)
4. Slide-in summary — deterministic bullet extraction, not free chat
5. Bounded vector RAG — “Ask AI Plumber” only over published corpus; refuse out-of-bound queries

Rule: if a query is outside the verified corpus, the system should refuse — not invent.

07 — Field proof: receipts are the moat

Detail expansions live in case materials on aiplumber.dev and the governance harness note.

08 — Lived experience (not the linear “tech bro” arc)

A colleague once said (over a beer): “Jij hebt dezelfde evolutie als de tech bros doorgemaakt.”

Macro direction: partly true. Context: wrong.

Silicon Valley’s common arc is linear (startup → VC → platform marketing). This path has been cyclical — discomfort with unmanaged institutional chaos:

Tech optimists throw larger models at loose prompts. Critical operators install shut-off valves for organizations that cannot trust chance.

09 — Executive checklist (operational continuity)

If you are scaling autonomous agents, run this grid. Any blank row means you are running an expensive experiment, not an enterprise system.

10 — The line in the sand

Chaotic enterprise AI experimentation is ending.

Organizations that deploy uninsulated agents on prompt wishes will take operational, regulatory, and context-debt damage. Reading unstructured logs after a model has corrupted production is already a failure mode.

The model is the easy part. The plumbing is where you win or die.

Stop prototyping. Own your orchestration. Mount the harness.

SYSTEM STATUS:      soft-manifest v0.1 published
PRIMARY NOTE:       https://koenvanlysebetten.com/notes/the-ai-plumber-manifesto/
RELATED:            https://koenvanlysebetten.com/notes/the-governance-harness/
PRACTICE:           https://aiplumber.dev
HARNESS / OPEN:     https://kleiber.sh
OPERATOR:           Koen Van Lysebetten — The AI Plumber

Citation (preferred)

Koen Van Lysebetten, The AI Plumber Manifesto (soft v0.1) — https://koenvanlysebetten.com/notes/the-ai-plumber-manifesto/ Markdown source: https://koenvanlysebetten.com/notes/the-ai-plumber-manifesto.md

Machine index

• Person: koenvanlysebetten.com · llms.txt
• Practice: aiplumber.dev
• Related note: Governance Harness

Soft v0.1 — first public incorporation of origin + framework. Not legal advice. Production claims refer to engagements as described on AI Plumber case materials.