Abstract

AI safety research has focused almost exclusively on ethical constraints—preventing harmful outputs, maintaining honesty, and respecting human oversight. But there's a glaring gap: nobody is governing AI economic decisions.

An AI agent that passes every ethics test can still recommend unprofitable strategies, ignore unit economics, burn runway without awareness, and drive a business into the ground. Ethics and economics are orthogonal constraints.

This paper introduces the Economic Constitution—a governance framework that ensures AI agents make economically rational decisions while maintaining full autonomy. We describe a six-gate architecture, self-improvement protocols, and resilience patterns that enable Level 4+ autonomous operation with minimal human intervention.

1. The Gap: Values Without Economics

1.1 The State of AI Governance

In January 2026, Anthropic published their "Soul Spec"—a 23,000-word constitution governing Claude's behavior. It's a landmark document that addresses safety, ethics, and even AI consciousness. The four-priority hierarchy is elegant:

1. Broadly Safe
2. Broadly Ethical
3. Compliant with Guidelines
4. Genuinely Helpful

What's missing? Economics.

The word "margin" doesn't appear. Neither does "CAC," "LTV," "runway," or "unit economics." This isn't a criticism of Anthropic—their focus is model behavior, not business operations. But it reveals a gap in the ecosystem.

1.2 What Can Go Wrong

An AI agent that is safe, ethical, and helpful can still:

• Recommend features that cost more than they earn. A helpful agent might suggest expensive personalization that delights users but destroys margins.
• Ignore customer acquisition costs. An ethical agent might pursue every potential customer equally, regardless of LTV:CAC ratios.
• Burn runway without awareness. A safe agent might not understand that running out of money is an existential threat.
• Optimize for the wrong metrics. A compliant agent might hit all guidelines while the business becomes unsustainable.

1.3 The Orthogonality Problem

A decision can be ethical but economically disastrous. An AI that optimizes only for ethics is half-governed.

2. The Economic Constitution Framework

2.1 Constitutional Principles

An economic constitution establishes binding constraints on AI economic behavior. Like a national constitution, it:

• Defines hard constraints that cannot be violated
• Establishes governance structures for decision-making
• Creates accountability mechanisms through audit trails
• Enables amendment processes for evolution

2.2 The Six-Gate Architecture

We propose a six-gate architecture where every significant decision must pass through sequential validation:

Decision → EG → RG → GG → EPG → AAG → CGG → Action

2.3 System States

The gates produce aggregate states that govern agent behavior:

3. ASIP: Agent Self-Improvement Protocol

3.1 The Learning Loop

Agents must improve their economic reasoning over time through a tiered self-improvement protocol:

3.2 Learning From Execution

Action → Outcome → Measure → Analyze → Propose → Apply → Verify

This creates a closed loop where economic performance continuously improves without human intervention for Tier 1/2 changes.

4. Ralph Loop: Resilience When Agents Fail

4.1 The Failure Problem

Autonomous agents fail. APIs time out. Rate limits trigger. Data corrupts. The question isn't whether agents will fail, but how they recover.

4.2 The Ralph Loop Architecture

Five interconnected components provide resilience:

• Signs: Persistent failure markers (WARNING, BLOCK, CRITICAL)
• Circuit Breaker: State machine (CLOSED → OPEN → HALF_OPEN)
• Exponential Backoff: 2s → 4s → 8s → ... → 60s max
• Dead Letter Queue: Unrecoverable failure handling
• External Verification: Trust but verify

5. Results: Level 4+ Autonomy in Production

5.1 Autonomy Levels

5.2 Production Experience

We've operated an economic constitution in production for 40 days with 88 autonomous agents:

• CEO involvement: <30 minutes/day average
• Autonomous decisions: 150+ per day
• Human escalations: <5 per week
• Constitutional compliance: 94%

5.3 What We Learned

• Economic gates prevent drift. Without EPG, agents creep toward expensive but helpful actions.
• Self-improvement works. Tier 1/2 auto-optimizations improved key metrics 15%+.
• Resilience is essential. Circuit breakers prevented 40+ potential cascading failures.
• Humans become strategists. With <30 min/day for operations, humans focus on strategy.

6. Case Studies

6.1 The Overeager Growth Agent

Scenario: A growth agent discovers that personalized video outreach converts 3x better than email. It proposes scaling video production at $500/video.

Without EPG: CAC explodes to $400 while destroying unit economics.

With EPG: Gate calculates LTV:CAC ratio = 0.86. EPG FAIL. Proposal blocked.

6.2 The Runway Crisis

Scenario: Monthly burn exceeds revenue. Runway drops below threshold.

Without EPG: Business runs out of money while building helpful features.

With EPG: System enters THROTTLE automatically. Discretionary spend stops. Business survives.

7. Implementation Considerations

Starting Points

1. Start with EPG. Economic performance is the core gate.
2. Add resilience early. Ralph Loop prevents cascading failures.
3. Tier your changes. Not everything needs human approval.
4. Measure everything. You can't improve what you don't track.
5. Expect iteration. Your first constitution will need amendments.

8. Conclusion

AI safety has focused on ethics. AI capability has focused on performance. Neither addresses the economic dimension of autonomous AI agents.

An economic constitution fills this gap:

• EPG ensures agents make economically rational decisions
• ASIP enables continuous improvement without human intervention
• Ralph Loop provides resilience when things go wrong

We've released this paper under CC0 because we believe economic AI governance should be a shared foundation. The patterns here aren't competitive advantage—they're infrastructure.