GADR: Gathering Architecture Decision Records from Meeting Transcriptions
This paper introduces GADR, a multi-agent, self-correcting workflow that effectively extracts architectural decisions from raw, noisy meeting transcriptions to generate structured Nygard-formatted Architecture Decision Records (ADRs), demonstrating superior stability and utility compared to single-pass LLM baselines while highlighting critical trade-offs between RAG-based enrichment depth and transcript faithfulness.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Software architecture is the invisible blueprint that holds a digital system together, determining how its parts talk to each other and where they live. But a blueprint is useless if the people who draw it forget why they chose a specific design or how they solved a problem. In the past, teams tried to fix this by writing down every major choice in a document called an Architecture Decision Record. These records are meant to be simple notes that explain what was decided, why it was chosen, and what alternatives were rejected, ensuring that future developers can understand the history of the project. However, creating these records is hard work. In the real world, important decisions are rarely made in neat, written reports. Instead, they happen in messy, unstructured meetings where people talk over one another, wander off-topic, and leave their reasoning buried in casual conversation. Extracting a clear, formal record from such a chaotic conversation has long been a barrier to keeping software projects organized.
A team of researchers at the Federal University of Pernambuco in Brazil has proposed a new way to bridge this gap using artificial intelligence. They built a system called GADR, which stands for Gathering Architecture Decision Records. The core idea is to take the raw, noisy transcript of a meeting—where people might be speaking in Portuguese, interrupting each other, or discussing unrelated topics—and automatically turn it into a structured, readable decision record. The researchers did not simply ask a computer to summarize the text. Instead, they designed a workflow where multiple specialized software agents work together in a loop. One agent listens to the conversation to find potential decisions, another acts as a critic to check if those decisions make sense or if the computer is making things up, and a third agent gathers extra technical information to flesh out the details. This process repeats, refining the output until it is ready for a human to review.
To test if this approach actually works, the researchers gathered transcripts from five real software development meetings. These included sessions with undergraduate students building video games and logistics platforms, as well as a meeting with senior researchers working on a health technology project. They fed these raw transcripts into their new multi-agent system and compared the results against two simpler methods: asking an artificial intelligence to write a record in one go without examples, and asking it to write one after showing it a few examples of good records. The study found that the new system was much more stable. The simpler methods often produced inconsistent results, sometimes merging unrelated ideas or missing key points entirely. In contrast, the multi-agent system consistently produced records that followed a standard format and captured the main decisions discussed.
When the researchers asked the students who participated in the meetings to review the generated records, the response was largely positive. Out of 55 specific evaluations of the decisions, 52 agreed that the system had correctly identified what was discussed. Furthermore, 51 of those evaluations stated that the generated records were clear and complete enough to be useful. The students reported that reading the records helped them remember the reasoning behind their choices, which is often the hardest part of software development to recall weeks or months later. Senior architects also reviewed the work and found that the system successfully identified about 18 out of 23 major decisions that the experts had spotted in the same transcripts. This suggests the system can find the most important architectural choices even in a sea of casual conversation.
However, the researchers were careful to note that the system is not perfect and should not be trusted to write the final version of a document without human oversight. The system sometimes added details that were not in the original meeting. For instance, in one case involving a video game, the system included a specific number of expected users that was never mentioned by the students; it had pulled this number from a different project's records while searching for context. This highlights a trade-off: the system makes the records richer and more educational by adding technical context, but it risks including information that was never actually agreed upon by the team. The researchers concluded that the best use of this tool is as a draft generator. It can turn a chaotic meeting into a structured first draft that a human can quickly check, correct, and approve.
The study suggests that the biggest hurdle to keeping software documentation up to date is not the lack of good ideas, but the effort required to write them down. By automating the initial extraction of decisions from messy conversations, this approach lowers that barrier. It does not replace the need for human judgment, but it shifts the work from writing from scratch to reviewing and refining. For the students in the study, this meant they could learn from their own discussions more effectively, seeing their informal thoughts transformed into professional documentation. The researchers believe this shift, from creating documentation to reviewing it, is a significant step forward for how teams manage the knowledge that builds their software.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.