Structured context cuts agent cost, but faster review still carries quality risk

Three systems improve agent work by narrowing evidence before generation. E3 estimates task scope and expands only after failed verification; on 121 controlled edits, it retained 100% success while cutting cost by 85%, tokens by 91%, and inspected files by 92%. Harness Handbook maps requested behavior to source locations, improving planning win rates while reducing planner tokens. CT-Repair compresses static and runtime evidence into queryable graphs; its filters reduced candidate methods by 94.85% before multi-perspective diagnosis. These results support selective disclosure over indiscriminate long-context loading, although the evaluations cover specific repositories and benchmarks.

Generation is most credible when its target behavior can be checked. Monty filters candidate formal specifications with syntax, fuzz testing, and clause-level conformance; precision rose from 75% to 91.6% in one dataset and from 64% to 85% in another. Use-case-oriented regeneration similarly replaces only dependency behavior exercised by a repository’s checks. It preserved 99.8% of observed validation behavior across 180 repository–dependency pairs and reduced exported API surface by 93.1%. The latter does not prove full semantic equivalence: 14 attempts failed, especially around edge cases and deep framework integration.

Measured productivity gains now span both implementation and review, but their quality signals diverge. In a controlled study of 49 developers, design-system-aware AI reduced completion time by 46.7% to 69.4% across Angular, iOS, and Android, while task completeness reached 96%. By contrast, an observational analysis of 1.02 million pull requests found agent-involved reviews were often faster, yet review-smell prevalence generally increased. The studies measure different tasks and outcomes, but together they argue against treating throughput as a sufficient deployment metric.