FIELD NOTE / LINKEDIN
Who writes the agent's exam?
The short film, the complete written thought, and the evidence behind it.
The LinkedIn edition will be linked here after its public post is verified.
Who writes the agent's exam?
Day 01 · 2026-09-28 · LinkedIn
Short video caption
An agent can write implementation and tests from the same mistaken assumption. Keep consequential acceptance contracts and fault fixtures outside its writable worktree. CI checks the expected revision, nonzero tests, fresh reports, independent provider outcome and a killed mutant. A green agent transcript is evidence, not signoff. Checkout case is illustrative; research on agent-generated tests motivates scrutiny, not a universal failure-rate claim. #EricFieldNotes
Full written post / accessible read
Let an agent implement a feature and write its own tests. That is useful speed. But if both are built from the same mistaken assumption, the suite can pass while the business contract is missed. The issue is who controls the acceptance test.
In a checkout review, the patch and its test may both assert that a confirmation screen appeared. That says little about the provider ledger after an ambiguous retry. This is an illustrative failure. Recent research also found that agent-written tests often use observational checks, so test count alone deserves scrutiny.
Write the consequential invariant before implementation. Keep that contract and the provider test fixture outside the agent's writable worktree. Let the agent add useful unit tests, but CI injects a lost acknowledgment after capture and reads provider state independently. A deliberate duplicate-key mutant must make the contract fail.
Give each run a worktree and immutable contract version. CI checks the expected revision, nonzero relevant tests, fresh reports, independent outcome and a killed deliberate fault. Block merge when any piece is missing. Do this because a fluent green transcript is only a witness; a protected test with a negative control earns the verdict.
#EricFieldNotes
Runnable illustrative contract: commerce harness
Evidence and boundary
On-screen boundary: RESEARCH + ILLUSTRATIVE GRADER. The sources below support documented mechanisms and specifications; illustrative scenarios are not presented as measured incidents.