Article: Agentic Fitness Functions: Extending Evolutionary Architecture Beyond Deterministic Rules

Details

Evolutionary architecture relies on fitness functions to enforce architectural intent continuously. The authors argue that deterministic checks (dependency direction, latency budgets, contract shapes) should remain the primary enforcement mechanism, but they fail for judgment-heavy concerns: does a service interaction respect domain boundaries semantically? Has an architectural decision record become stale relative to current system state? The article introduces 'agentic fitness functions' for these gaps. An agent is scoped to the artifact changed, presented with structured evidence (AST diffs, ADR documents, semantic contracts), and required to return a structured verdict with confidence score. Low-confidence or high-blast-radius outcomes escalate to humans rather than auto-blocking. A key architectural requirement: the agentic signals are advisory, not blocking gates. They run alongside deterministic checks. The long-term payoff is that recurring agent findings can be converted back into deterministic rules, making architectural judgment observable and calibratable. The article comes from an InfoQ Certified Architect Program cohort and is thorough in practical implementation details, including versioned rubrics, evidence scoping, and escalation policies.