The team and the problem
The client is a small policy research team in Washington, DC. Their days are full of the work that surrounds research rather than the research itself: checking what the organization published, chasing open items, writing concept notes and agendas, finding background on experts, and translating documents for colleagues and partners.
None of those tasks is hard. Together they take hours every day, and they fall unevenly on whoever is most available. The team wanted one assistant that everyone could use, that knew the team's context, and that did not require anyone to learn a new tool.
Where the agent lives
The agent works in WhatsApp. Team members message it the same way they would message a colleague. There is no dashboard to open and no login to remember.
That choice shaped everything else. Because the agent knows who is writing, it can keep track of each person's items while still sharing one memory across the team.
A day in the life
The agent runs on a schedule in the morning and on request for the rest of the day.
- Each morning it checks for new publications and events in the team's field, logs them, and flags the notable ones.
- Team members can get a short list of their open items for the day.
- During the day, anyone on the team can ask it to draft, research, translate or record something, and the result comes back in the same chat.
What it keeps track of
The agent maintains shared logs that the whole team can read and add to. They cover decisions, deadlines and document status.
This is the part that makes it feel like a team member rather than a chatbot. When someone says a report is now final, that status is recorded for everyone. When the team agrees a format for a document, later drafts follow it.
It can also draw on a regularly updated knowledge base for the team's subject area, so background questions get answers grounded in current material rather than whatever a general model remembers.
What it produces
- Event concept notes in the team's own format.
- Agendas in the same format, so a draft is ready to edit rather than written from a blank page.
- Expert profiles built from public sources, for speakers, panelists and contacts.
- Translations of documents and messages in the languages the team works in.
- Messages to teammates. When asked to contact a colleague, it drafts the message, confirms it with the person who asked, and then sends it.
Guardrails
People approve anything the agent sends on their behalf. The draft, confirm, send pattern is deliberate. It keeps a person responsible for every outgoing message while removing the time it takes to write it.
It works inside the limits the team agreed, and the shared logs mean anyone can see what it recorded.
We do not publish time-saved figures for this engagement. The honest description is that the morning routine now happens before anyone logs on, and drafting and translation requests come back in the same chat.
How we built it
We build every agent in four steps: map the repetitive work that costs the team the most time, set the agent up with the team's documents, templates and rules, launch it with clear limits on what it can send, share and change, and keep improving it as the team uses it.
If your team spends its mornings on tracking, drafting and translation rather than on the work you were hired to do, this is the kind of agent we build.