AgentExecutor Runs Arbitrary Code Snippets at 94% Coverage While Cutting Execution Time 80.3% and Cost 56.6%

arXiv 2608.05959·medium signalXBlueskyLinkedInCopy link

AgentExecutor (arXiv 2608.05959, Aug 6) tackles partial code execution — running a snippet without its missing context and dependencies — with a three-phase multi-agent design: environment preparation, dynamic exploration with iterative refinement, and prefix evolution via program synthesis. Unlike LExecutor and Treefix it can create resource files and resolve environment configuration, and it adds coverage-guided context pruning. On Stack Overflow snippets and open-source project code it reaches 94% and 90% coverage, beating Treefix by 19.9% and 13.8%, while cutting execution time up to 80.3% and cost up to 56.6%. Practical for anyone building dynamic analysis or sandboxed snippet evaluation into a tool.