Inspiration
Small problems in the workplace are often noticed first by the people doing the work. However, these concerns may remain unresolved because they are difficult to organize, explain, or turn into a realistic action.
We built NEXT ACTION AGENT to help workers transform vague concerns into one small, practical first step—while keeping the final decision in human hands.
What it does
NEXT ACTION AGENT guides the user through a simple improvement workflow:
- The user describes a workplace concern in their own words.
- Gemini organizes the concern into a neutral summary.
- The agent asks only one or two necessary follow-up questions.
- It generates three possible improvement ideas.
- It compares them using cost, time, safety, authority, and reversibility.
- It recommends a small first action.
- The improvement card is saved to Firestore.
- The user can later record the result and receive a suggested next step.
The goal is not to make decisions for the user, but to reduce uncertainty and make continuous improvement easier to begin.
How we built it
The backend was built with Python and FastAPI. We used the Google Gen AI SDK to connect Gemini through Vertex AI.
The agent is deployed as a containerized service on Google Cloud Run in the asia-northeast1 region. Firestore stores improvement cards, user answers, selected actions, review dates, and feedback results.
The frontend was built with HTML, CSS, and JavaScript and is served by the FastAPI application. The complete application flow runs on Google Cloud:
- Gemini on Vertex AI for reasoning and structured suggestions
- Cloud Run for hosting and automatic scaling
- Firestore for persistent improvement records
- Docker for reproducible deployment
Challenges we ran into
One challenge was converting unstructured workplace concerns into consistent, structured responses from Gemini.
We also needed to limit follow-up questions so that the agent could understand the situation without overwhelming the user. Designing recommendations around safety, cost, authority, and reversibility required careful prompt design.
Another challenge was creating a feedback loop rather than a one-time AI answer. We connected the generated improvement card with Firestore so users can return later, record what happened, and continue with another small step.
Finally, we worked through Cloud Run deployment, permissions, environment variables, and Google Cloud service configuration to make the complete application publicly accessible.
Accomplishments that we're proud of
We completed an end-to-end agent workflow that turns a vague concern into a concrete and manageable first action.
The application uses Gemini for real reasoning, stores improvement history in Firestore, and runs as a live Cloud Run service. It also supports follow-up reflection, allowing the user to record whether an action worked and receive a suggested next step.
We are also proud that the interface keeps the user in control. The agent provides options and reasoning, but the final judgment always belongs to the person doing the work.
What we learned
We learned that the value of an AI agent is not simply generating more ideas. Its real value is reducing the effort required to understand a problem, compare realistic options, and decide where to begin.
We also learned that small and reversible actions are especially important in workplace improvement. Recommendations become more useful when they consider the user's authority, available time, safety, and ability to return to the previous state.
Finally, persistent records turn AI advice into an ongoing improvement process rather than a one-time conversation.
What's next for NEXT ACTION AGENT
Next, we plan to add:
- User authentication and team workspaces
- Multilingual support
- Review-date reminders and notifications
- Team sharing and collaborative improvement cards
- Organization-specific safety rules and operating procedures
- Analytics for recurring workplace concerns and improvement outcomes
- Evaluation tools to measure the quality and usefulness of agent recommendations
Our long-term goal is to make continuous improvement accessible to every worker—one small, safe, and practical action at a time.
Log in or sign up for Devpost to join the conversation.