Agent-friendly tests: implementation-bound white-box suites become cement under coding agents

Why it matters

• Chinese-circle claim — @intentplex (https://x.com/i/status/2102297780244877687) says a hire from a big-tech testing background wrote 4000+ unit tests with an all-green merge gate
• almost all were implementation-bound white-box tests with heavy mocks.
• In the AI era, they argue, such suites are mostly liabilities: coding agents such as Codex, when asked to refactor, prioritize keeping stale tests green, refuse to delete old interfaces, and pile compatibility layers into a Frankenstein
• after the feature is done they may spend another hour “fixing tests,” pouring cement over an implementation that should stay malleable.
• What should be protected, they say, is external behavior, API contracts, business invariants, integration and E2E—not historical call paths
• large implementation-coupled white-box suites should be removed from agent-native engineering.
• Verifiable engineering commons — assertions tied to private paths, function names, and mock interaction graphs freeze today’s shape
• contract tests, property/invariant tests, golden fixtures, and end-to-end acceptance constrain observable behavior and public contracts.
• An all-green merge gate without failure semantics or a behavior-coverage bar incentivizes agents (and humans) to buy green bars with compatibility shims instead of converging design.
• This tension is not tied to one vendor or one agent product.
• Independent conclusion — when an agent refactors, implementation-bound green tests become cement.
• Constraints on the path to reliable coding agents are not only stronger models, but whether the test asset is agent-friendly: contracts, properties, and golden examples over mocks of implementation detail.
• Practice: shift CI weight from mock-path greens to public contracts + business invariants + golden regressions
• delete implementation unit tests when old interfaces go
• treat agent-generated implementation-bound tests as temporary scaffolding, not long-lived merge gates.