Coding agents improve when specifications become executable and reusable

ReProAgent converts issue reports into fail-to-pass tests through repository exploration, code-graph retrieval, and runtime validation. It reproduced 58.43% of SWT-bench-lite issues and 70.30% of SWT-bench-verified issues at an average cost of $0.14 each. TestAgent applies a planner, generator, and reviewer around a test-specific repository graph. On six Java projects, it reached 92.34% line coverage and an 83.69% mutation score; it also detected 154 real bugs with 92.22% precision.

The input report matters as much as the execution loop. An analysis of 441 SWE-bench Verified bug reports found that affected-code locations and suggested fixes had the strongest positive associations with successful repair. This supports concise reports that narrow the search and repair space.

DualVeri uses shared property templates for Lean 4 proofs and property-based testing against PySpark. Across 400 candidate properties, templates raised proof-synthesis success by 1.6× on average, cut proof hallucinations by 59%, and reduced test-intent mismatches from 22 to 1. Disagreements between proof and execution exposed gaps between the formal model and runtime behavior.

Diversify2Verify adds another useful control: generate several implementations under fixed contracts. Array, list, recursive, and imperative variants produced at least one verified artifact for 49 of 73 tasks, a 67.1% task success rate. Verification success therefore depends partly on implementation structure, even when variants target the same task.

Selective persistent memory stores task specifications, data schemas, tool settings, and output constraints while excluding temporary traces. On 24 recurring enterprise tasks, it achieved 96% completion, compared with 79% for stateless sessions and 71% for full-history memory. Compatible data refreshes completed all 18 tasks without another LLM call.

Public skill repositories show that reuse remains uneven. Among 11,497 software-engineering skills, implementation, testing, and code review accounted for 65.4%. Only 13.6% included executable code assets, and requirements, release, and deployment together accounted for 10.7%. Most published skills still package instructions, leaving substantial room for tested, executable components.