A new multi-agent AI system wants to write your software requirements documents for you.
Researchers published Agent4RE, a framework that splits the job of requirements engineering - turning vague stakeholder wishes into formal specs - across multiple specialized AI agents instead of a single prompt. The system runs two feedback loops: one where agents critique and revise their own output, and one where a human can step in and correct it. To test it, the team built a new dataset called RE-E2E, drawn from real, human-written requirements documents, since prior benchmarks only covered isolated pieces of the process like extraction or classification. They ran the system across eight different large language models and compared results against a baseline that simply stuffs the prompt with domain context.
Requirements engineering is the unglamorous, expensive part of software projects - figuring out what to build before anyone writes code - and it resists automation because it involves ambiguity, negotiation, and judgment calls that don't fit neat prompts. The 8% average improvement over the baseline is real but modest, and the larger gain, about 0.8 points on a four-point human and judge rating scale, came specifically from the versions with self-refinement or human feedback loops, not from bigger models alone. That suggests the actual unlock is process - splitting the task into agent roles and adding iteration - rather than any raw capability jump.
A four-point rating scale invites skepticism about precision, but if the results hold up outside the lab, it's one more sign the next wave of AI tooling is aimed at the paperwork around code, not just the code itself.