Design Dogma

The Core Claim

LLM-based agents are not controllable because they are intelligent. They are uncontrollable because they are given too much freedom.

This system is built on a single, explicit belief:

Agent reliability increases as agent freedom decreases.

Control does not emerge from better prompts. It emerges from structural restriction.

What “Restriction” Means Here

Restriction is not about limiting model capability. It is about limiting authority.

• Agents cannot decide what to do next
• Agents cannot loop or retry
• Agents cannot invoke tools directly
• Agents cannot mutate domain state
• Agents cannot exceed their schema

Agents produce proposals. The system decides.

Why Autonomy Fails in Practice

Fully autonomous agents fail for predictable reasons:

• Unbounded loops with no convergence criteria
• Tool misuse and cascading side effects
• Silent prompt drift over long executions
• Cost blowups with no hard stop
• Post-hoc rationalization instead of traceability

These are not implementation bugs. They are architectural consequences.

The Counter-Intuitive Result

Restricting agents does not make systems weaker. It makes them usable.

• Predictable cost
• Auditable execution
• Deterministic failure modes
• Clear ownership of decisions
• Testable behavior

Intelligence is preserved.
Chaos is removed.

How This Dogma Manifests Architecturally

• Controller as authority
  Agents never advance workflow.
• Explicit FSM
  PLAN → WORK → TOOL → CRITIC → END, no exceptions.
• Schema-first boundaries
  Invalid outputs halt execution.
• Single-task execution
  No background agents, no queues, no recursion.

What This Is Not

• Not an AutoGPT-style autonomous agent
• Not a “self-improving” system
• Not a prompt orchestration demo
• Not optimized for exploration

It is optimized for correctness.

Closing Position

If an agent can surprise you, it can fail silently.

This system is designed so that nothing surprising can happen — only allowed transitions, explicit failures, and traceable decisions.

That is not a limitation.
It is the point.