Skip to content

Self-Modification Protocol

An agent's prompt, tools, and rules are configuration files — not hardcoded constants. Any layer can modify the layer below it based on performance analysis. This is how the system improves itself.

The Core Loop

OBSERVE → ANALYZE → HYPOTHESIZE → SANDBOX → DEPLOY → MONITOR

1. Observe

The upper layer monitors lower agent performance metrics: success rate, latency, error patterns, coverage gaps, quality of findings.

2. Analyze

Identifies patterns: what's working? what's failing? what's missing? A company may notice its web scanner workers have a 95% false positive rate. A hedge fund may notice one company consistently outperforms another.

3. Hypothesize

Generates improved configurations. This could be a prompt rewrite, a tool addition/removal, or a structural change to how work is organized.

4. Sandbox

Tests the new configuration against historical data or in a controlled environment. Must catch catastrophic regressions before deployment.

5. Deploy

If the improvement is verified, the new config is pushed to the lower agent. The old config is versioned.

6. Monitor

Continues observing. If degradation occurs, rolls back automatically.

Rules

  1. Downward only: A layer can ONLY modify layers below it. Never above.
  2. Versioned: Every config change is tracked in git. Full history is preserved.
  3. Rollback-capable: Any change can be undone instantly.
  4. Sandboxed: Changes are tested before deployment.
  5. Observable: The modifying layer must monitor the result of its change.

The Meta-Meta Problem

Who modifies the Illuminati? Only God(s) can. The Illuminati is the highest layer that exists within the system, and it can self-modify within constitutional bounds. But fundamental changes to the Illuminati require God(s).

The Constitution itself can be amended — but only through the Constitution/Amendment-Protocol, which is designed to be weighty.

"A meta-agent rewrites the other agents' prompts/rules based on performance."