AI agents struggle to deliver complex multi-step professional projects
DeliverBench: Benchmarking AI Agents on Delivering Realistic Multi-Stage Professional Projects
Artificial Intelligence
Summary
Many AI systems are good at finishing small, individual tasks, but in real jobs, success means completing an entire project correctly from start to finish. The authors created DeliverBench, a new test that checks if AI agents can handle full projects with many steps and changes along the way. They found that even the best AI systems only succeed about half the time on these complex projects, showing that finishing projects is harder than just doing separate tasks well. This means AI needs to get better at keeping track of and improving work over time.
What this means in practice
- •For enterprise ai teams: Test and improve AI systems for handling end-to-end professional workflows with multiple feedback rounds and evolving inputs.
- •For legal service providers: Evaluate AI tools for drafting and refining complex legal contracts across multiple stages and client revisions.
Authors
Yumou Liu, Mingju Chen, Zhoufutu Wen, Qianyu Yang, Zijie Wang, Yeqian Liu, Chenghao Yang, Tiliang Duan, Qiyu Cai, Yanglihong Xiao, Jiayun Dong, Jianlan Ma, Shenhao Miao, Yuhan Cao, Lu Xu, Que Shen, Tian Gao, Ge Zhang, Jianpeng Jiao, Xuanhe Zhou, Fan Wu
Abstract
LLM agents are judged by what they deliver, and in professional work what gets delivered is a project, not a task. A project pays off only if the whole package holds, and one weak part can limit the value of the rest. Yet existing benchmarks mostly test agents on one task at a time, leaving a gap between the capability we measure and the outcome that makes agents useful. To address this gap, we introduce DeliverBench, which tests whether agents can carry a project from brief to final delivery. Its 11 projects, spanning financial due diligence, legal contract drafting, and drug safety reporting, are reconstructed with 13 domain experts from real materials and chain 150 subtasks through 2-15 rounds of new materials and client feedback. Because no reference outputs are injected, agents inherit their own errors from one task to the next. Expert rubrics then score every checkpoint as well as the final package. Across seven model-harness systems, even the best system scores only 65.0 out of 100 on final delivery and passes 50.0% of subtasks at a 90-point threshold. The failures are often failures of carrying work forward: in one audit project, all seven systems addressed the review findings, but only one preserved traceability. Together, these results suggest that task-level capability does not simply compose into delivery, which also depends on keeping the work coherent as it evolves.