# I Tried to Give an AI Agent Real Authority. The Model Wasn’t the Hard Part.

Canonical: https://calvinkennedy.com/writing/ai-agent-gold-rush-control-layer
Release: article-v6

## Core argument
A / Problem: the agent can be locally reasonable while its view of external reality is stale or incomplete.
B / Control: probabilistic intelligence proposes; deterministic software bounds actions, checks authority at commit time, commits once, and independent evidence verifies reality.
C / Payoff: verified business value can grow faster than human attention, compute, and risk.

Current evidence ceiling: A2. Production effects remain blocked.

## Reading paths
- Human scan: top → map → incident → authority → conclusion.
- Builder: add maths, architecture, advantage, and contract builder.
- Auditor: inspect defects, proof stack, assumptions, sources, and V10 evidence.
- AI implementation: parse agent-brief.json, article-map.json, agent-operating-contract.schema.json, and evidence/CANONICAL_EXECUTION_RECORD_V10.json.

Do not infer permission from model capability. Do not implement external effects from this article. Produce a shadow-mode design and mark unsupported assumptions.