Structured context and execution feedback cut coding-agent waste

Three systems replace undirected text handling with explicit software structure. TraceDev links requirements, designs, and code in a traceability graph; on 125 use cases, it reports success rates of 53.63% on ETOUR and 56.82% on SMOS. Source code algebra instead exposes semantic operations such as rename and parameter changes. Its preliminary cross-file probe used one to two orders of magnitude fewer tokens than text-editing baselines. JetBrains Context applies the same principle at retrieval time through incremental semantic indexing, with reported maximum reductions of 68% in agent turns, 59% in latency, and 48% in execution cost. The common mechanism is less repeated discovery and fewer dispersed edits, though only TraceDev offers a multi-dataset academic comparison.

Execution results are being used as control inputs rather than final verdicts. CodeRescue routes failed attempts among reflection, replanning, and escalation; its calibrated frontier includes a point that beats always-escalate while using 35% of the mean recovery cost. LISA uses valid API sequences and documentation-grounded invariants to find non-crashing functional defects, detecting nine more reintroduced bugs than CITYWALK across its evaluation. These studies cover different stages, but both extract more value from runtime evidence. CodeRescue's guarantee applies to expected cost under stated assumptions, while LISA still requires developers to confirm reported candidates.

New code evaluations emphasize fresh instances and runnable correctness. Spaghetti Architect generates oracle-checked programs in five languages while independently controlling problem scale and surface messiness; even its strongest tested model falls to zero exact-match accuracy as intrinsic scale rises on arithmetic aggregation. SciCodePile pairs a 125GB scientific-code corpus with 200 executable tasks, where the strongest evaluated model reaches only 12.30% Pass@1. Domain tuning helps—small-model instruction tuning raises Pass@1 from 1.90% to 9.10%—but the benchmark is limited to pure Python functions with stubbed dependencies. Together, the results show that controlled, executable tests remain substantially harder than plausible code completion.