DeliverBench evaluates AI agents on multi-stage professional projects with cumulative errors

What happened.AI researchers introduced DeliverBench, an evaluation benchmark measuring how language model agents manage compounding errors and evolving requirements across multi-stage professional projects.
Why it matters.Real-world knowledge work involves iterative client adjustments that compound early mistakes over multi-week delivery cycles. Disabling artificial reference resets exposes how current language models struggle with long-term traceability and error recovery in production workflows.
The technical specifications:
- Professional projects: Features 11 realistic assignments spanning financial due diligence, legal contract drafting, and clinical drug safety assessments.
- Iterative feedback: Chains 150 subtasks through 2 to 15 rounds of revised client inputs and requirement shifts co-designed with 13 domain experts.
- Cumulative mistakes: Removes intermediate ground-truth corrections, forcing agents to carry forward and debug their own earlier mistakes across project milestones.
- Performance ceiling: Top-performing frontier agents scored 65 out of 100 on final deliverables, passing only half of subtasks under strict evaluation rubrics.
- Coordination breakdown: Shows that single-turn coding proficiency fails to predict project delivery success due to context loss during requirement iterations.
STORYLINE PROGRESSION & CONTEXT
This report represents a verified evolution of prior newsroom intelligence: