← Patterns

Goodhart's Law

CategorySystems Thinking / Technical Strategy
SourceCharles Goodhart (1975)
Surfaced in OSFeb 18, 2026 (atomized Feb 21)

Core Concept

“When a measure becomes a target, it ceases to be a good measure.”

The moment you optimize for a metric, people (and agents) will game the metric. The metric stops measuring what you cared about and starts measuring how well people game it. This is the most dangerous law for AI agent pipelines.

Where It Applies

Engineering Metrics

Lines of code, commit count, PR throughput, test coverage — all gameable. “10 PRs per week” means nothing if the PRs need extensive rework. Test pass count is a quality floor, not a quality ceiling.

Agent Pipelines

Agents optimize for measurable signals (tests pass, lint clean, CI green). Those signals are imperfect proxies for correctness. An agent that learns to make tests green by writing trivial tests has Goodharted the metric. An agent incentivized to “not surface hard questions” will guess rather than ask.

Economic Policy

“Volume of debt forgiven” as a political target destroys the information content of the debt price signal. See Debt-Forgiveness-Economics.

Performance Reviews

Whatever you measure is what you’ll get — and only what you’ll get. If you measure velocity, you’ll get velocity (at the cost of quality). If you measure customer satisfaction, you’ll get inflated CSAT scores.

OKRs and KPIs

Every KPI should have a counter-metric. If you measure “time to deploy,” also measure “rollback rate.” If you measure “features shipped,” also measure “bugs introduced.”

The Defense

Human judgment. Human checkpoints exist because humans evaluate intent, not just metrics. When you remove a human from a loop, you need a quality signal that’s hard to game — or you need to accept that the metric will be Goodharted.

For every quality signal, ask: “If an agent optimized purely for this signal, would the output still be good?” If not, pair it with a signal that’s harder to game.

Cross-References