Fitness Function

A fitness function is a quantitative measure that defines what "healthy" means for a software system in terms an agent can evaluate, enabling automated assessment of system quality without human interpretation.

Details

The concept is borrowed from evolutionary architecture and applied to agent-driven operations. Fitness functions make quality attributes machine-readable: response time percentiles, error rates, dependency health, schema conformance, test coverage thresholds, and architectural constraint compliance. Each function encodes a specific aspect of system health as a computable check with a clear pass/fail or numeric result.

Fitness functions are a prerequisite for self-healing systems (agents returning a system to a known good state after degradation) and self-improving systems (agents evolving non-functional qualities like performance and reliability over time). Without machine-readable definitions of "healthy," agents cannot assess whether an intervention improved the system or degraded it. The latent knowledge problem - senior engineers' undocumented pattern-matching about what "healthy" looks like for a specific system - must be encoded into fitness functions before agents can replicate that judgment.

Fitness functions are distinct from evals, which measure model or system output quality (did the agent produce a good response?). Fitness functions measure the health of the system the agent operates on (is the software system in a good state?). They complement harness engineering by defining the target state that mechanical enforcement aims to maintain - linters and structural tests enforce invariants, while fitness functions assess whether the overall system remains within acceptable operational bounds. They also complement observability, which provides the raw signals (metrics, traces, logs) that fitness functions evaluate.

Examples

• An SLO-based fitness function that defines "healthy" as p99 latency under 200ms and error rate below 0.1%, evaluated by an agent querying a metrics API after each deployment.
• A structural fitness function that checks dependency directions between architectural layers (e.g., UI may depend on Service but not vice versa), run as part of a harness engineering pipeline.
• A data quality fitness function that validates schema conformance, referential integrity, and data freshness across pipeline outputs, enabling an agent to detect and remediate data degradation automatically.
• A security fitness function that verifies no public endpoints were added without authentication middleware, checked by an agent scanning route configurations after each change.