Coding-agent controls gained precision through targeted context and executable evidence

Two studies favor selective context over exhaustive exploration. ACQUIRE asks repository-specific questions before repair and improved Pass@1 on SWE-bench Verified by up to 4.4 percentage points. E3 estimates task scope, begins with the smallest viable execution path, and expands only after failed verification; on its controlled benchmark, it preserved 100% success while cutting tokens by 91%. Together, they support context acquisition as a budgeted intervention rather than a default repository-wide scan. E3’s strongest numbers come from a simulator, so its efficiency gains need broader confirmation.

The harness is now part of both the measurement and the safety claim. AgentCompass found that changing the harness altered benchmark scores and trajectory failures. A supply-chain study likewise showed the same model catching or installing an attack depending on its harness, while a deterministic pre-install check closed most observed gaps. Proof-or-Stop adds a stricter lifecycle boundary: an agent’s claim advances only when fresh evidence is authenticated and tied to the current source state. Its tests blocked false completion and tampered receipts, although evaluation covered one model family and a self-hosted corpus.

Executable checks are becoming narrower and more domain-complete. DiffTestGen directs tests toward changed functions and exposed behavioral differences in 78.2% of 463 pull requests. GapForge targets uncovered compiler regions, adding tens of thousands of covered lines over a reported baseline and finding 12 real failures. Alipay-PIBench applies the same logic to payment integration, testing end-to-end behavior, signature and notification handling, idempotency, refunds, and business-state consistency. These results show that a generic pass signal is insufficient when the meaningful failure surface is a code diff, a coverage gap, or a domain invariant.