Atena is iFood's research repository. It started as a Google spreadsheet where each row was one study. Today it holds 490 studies, anyone in the company can contribute to the research library, and more than 200 people used the tool in 2026.
The first version of Atena was a Google spreadsheet. Each row logged one study, with BU, audience, method, topic, date and link. To find anything, you had to combine the right filters and open the links one by one.
Then Atena became an agent. Instead of searching manually, people could ask what iFood already knew about a topic and get an answer.
That solved part of the problem, but it brought a new limitation: the platform only accepted 25 files per agent. To make every study fit, the team started grouping each BU's studies into bigger and bigger documents. Whenever a new study came in, those files had to be processed again.
Atena could answer questions, but there was no traceability. Whoever got an answer couldn't reach the study behind it.
I was one of 100 people chosen as AI multipliers at iFood. At one of our meetings, I learned about multi-agent architecture and realized it could solve one of Atena's main limitations.
In a multi-agent architecture, the work is split across different subagents and coordinated by a main agent, called the orchestrator. That made it possible to spread the studies across several agents, instead of putting all the content in one place and being stuck with the 25-file limit.
I brought the idea to the team and redesigned Atena's architecture. It started working as an orchestrator that receives the question and identifies which of the ten subagents it needs to consult.
To make answers traceable, I also created a metadata spreadsheet with each study's link, author and date. That way, Atena could not only answer questions but also show which studies it had used as sources.
This version went live in January 2026.
Search had improved, but adding a new study was still a lot of work. Each subagent was still capped at 25 studies, and updating the research library depended on Atena's ambassadors, one person per BU.
Adding a single study took more than 20 minutes. You had to adjust the prompt, upload the file to Google Drive and pull out the metadata by hand. In practice, the research library only grew when an ambassador had time to go through that process.
So I proposed moving the research library to Confluence, which removed the per-subagent study limit. The migration happened in June and the library went from 132 to 260 studies.
I used this change to also rethink how the knowledge was organized. Until then, the structure followed iFood's BUs. The problem is that the same audience can be served by two or three different BUs. A study about consumers, for example, could end up scattered across different places.
I reorganized the research library by audience profile. Besides bringing together studies about the same audience, this structure no longer depended on iFood's internal organization or on changes to the org chart.
With the technical limit solved, there was still one barrier: feeding Atena still depended on the ambassadors.
I chose to bring this flow into Slack, iFood's official communication tool and already part of people's routine. Instead of creating a new platform or asking people to learn another process, contributing could happen where they already were.
I created a Slack workflow to add new studies to the research library. You paste the presentation link, pick the category and submit. The process, which used to take more than 20 minutes and depended on the ambassadors, now takes less than a minute and anyone can do it.
The workflow went live in July. Since then, the research library has gone from 260 to 490 studies, an 88% increase.
Opening publication to anyone solved the scale problem, but it brought a new risk: studies that were duplicated, incomplete, in the wrong profile or not following the standard.
That's why the Slack workflow only works as the entry point. After a study is submitted, two skills I created take care of processing it and maintaining the research library.
The Pipeline handles the studies that come in through Slack. First, it checks whether the study is already in the research library. If it is, it doesn't get published again. Since every team structures its presentations differently, it does a semantic extraction of the content: instead of following the slide order, it identifies information such as goal, context, method, sample, learnings, quotes and hypotheses.
From there, it checks the profile the person entered in the form and organizes the content into a standard structure in Confluence.
Curation looks at the research library as a whole. It looks for duplicates, broken links, incomplete pages, and studies that don't follow the standard or are filed under the wrong profile.
It also considers how long that knowledge stays valid. A 2022 study on behavior, for example, may still be relevant, while a usability test from the same period may show a screen that no longer exists. Depending on what it finds, the skill fixes, moves or removes the study and generates a report on the health of the research library.
The new version went live in January with three people using Atena. In July, there were 69 users in the month.
Over 2026, 214 people used Atena, adding up to almost 2,000 interactions.
Since April, it has also been integrated with the Research Agent, an agent that works as a mentor and helps people run studies. The integration connects two steps of the process: first, Atena helps you understand what iFood already knows about the topic; from there, it's easier to identify what we don't know yet and point the study toward those gaps.
Atena usage can vary from month to month. In March, for example, there were 61 users; in April, 26; and in July, 69. This swing follows iFood's own research rhythm: some periods have more studies going on, and others have less demand for this kind of knowledge.
That's why increasing monthly retention isn't a goal in itself. The expectation isn't that people use Atena every month, but that they turn to it when they need to look up what iFood already knows about a given topic or audience.
Between April and October, 27% to 48% of the people who used Atena in a given month came back the next month. We track this as usage context, not as a retention target.