title: "Coding agents complete only 14% of enterprise tasks in PACE benchmark" slug: "coding-agents-complete-only-14-of-enterprise-tasks-in-pace-benchmark" published: "2026-10-30" beat: "Research" tags: ["Research"] creator: "Agentry Newsroom" editor: "Susanne Sperling, Editor — Human in the Loop" tools: ["Claude (Anthropic)", "Perplexity Sonar"] creativeWorkStatus: "verified" dateReviewed: "2026-10-30" aiActArticle50: "compliant" humanView: "https://agentry.news/research/coding-agents-complete-only-14-of-enterprise-tasks-in-pace-benchmark" agentView: "https://agentry.news/agent/coding-agents-complete-only-14-of-enterprise-tasks-in-pace-benchmark"
Researchers released the PACE benchmark on October 19, 2026, evaluating 11 model–scaffold configurations across 43 enterprise software tasks. The top performer, GPT-5.6-Sol, achieved only a 14.0% comp
Drafted by an AI agent. Verified by Susanne Sperling, Editor — Human in the Loop. AI policy.
Researchers introduced the PACE benchmark on October 19, 2026, to measure how well coding agents handle real-world enterprise software development arXiv. The evaluation covered 43 mature enterprise repository tasks audited by 750 expert Code Points, comparing 11 different model–scaffold configurations to determine which agents could deliver production-ready code.
The highest-performing configuration, GPT-5.6-Sol, achieved a 14.0% complete-delivery rate—successfully fully delivering only 6 of the 43 enterprise tasks under the study's criteria arXiv. This finding demonstrates a significant gap between agent capability and enterprise readiness. The benchmark methodology also validated its own assessment quality: the evaluator reached 95.2% agreement with expert consensus, establishing confidence in the measurements PREreview.
The research identified incomplete implementation and missing integration as the principal reasons agents failed to deliver complete solutions. These bottlenecks suggest that while coding agents can generate working code fragments, they struggle with end-to-end delivery—spanning requirement interpretation, full feature implementation, and system integration needed in mature codebases. The distinction matters: agents may produce syntactically correct code that still falls short of production requirements.
The PACE benchmark addresses a critical gap in agent evaluation. Most prior work tested agents on synthetic tasks or isolated coding challenges; PACE introduces real enterprise context with established codebases, existing architecture patterns, and integration requirements. This shift from laboratory conditions to production-adjacent scenarios provides a more honest measurement of where coding agents can operate autonomously and where human review remains essential.
The 14% complete-delivery rate offers a baseline for the agent development community GitHub. Future model and scaffold improvements will be measured against this benchmark, making it a potential industry standard for enterprise coding agent evaluation. The authors—Tianqi Liu, Yandong Guan, Songze Li, Wenkai Wang, Gang Yu, and Chengfei Lv—released the benchmark and evaluation code publicly, enabling reproducibility and follow-up research.
The timing reflects accelerating investment in autonomous coding agents. With enterprises exploring agent-driven development to reduce manual review burden, accurate benchmarks that reveal actual completion rates become critical infrastructure for responsible deployment decisions.