-
-
Arch Dgm
Inspiration
Family shopping sounds simple, but every household member can have different preferences, cuisines, allergies, health needs, and priorities. Most AI assistants answer individual prompts and then stop. We wanted to explore what happens when an AI takes a family goal and manages it as a persistent mission instead.
That idea became Fam Agent, an autonomous family shopping and errand coordinator.
What it does
Fam Agent maintains reusable family context, including dietary preferences, preferred cuisines, likes, dislikes, allergies, foods to avoid, and temporary health conditions.
A family can give it a natural-language goal such as:
Plan groceries for my family for this week under $150.
Fam Agent uses Gemini to understand the goal and existing family context, asks for clarification only when necessary, and transforms the request into a structured, persistent mission.
The current prototype demonstrates the mission lifecycle from understanding and planning through Monitoring.
How we built it
Fam Agent uses a single-agent architecture.
Gemini provides reasoning, while Google Agent Development Kit (ADK) provides agent orchestration and specialized tools.
The backend is built with FastAPI and deployed on Google Cloud Run. Firestore persists family profiles and mission state.
A dedicated mission runtime manages lifecycle, persistence, state transitions, and bounded execution around the ADK agent.
Architecture:
Family UI → FastAPI / Cloud Run → Mission Runtime → Google ADK Agent → Gemini + Tools → Firestore
Challenges we ran into
The biggest challenge was making Fam Agent an agent rather than a chatbot.
We had to carefully separate AI reasoning from deterministic application logic. Gemini decides how to interpret the goal, what information is missing, and how to construct the plan, while the mission runtime controls persistence and lifecycle.
We also needed family profiles to contain useful context without turning onboarding into a huge questionnaire.
Accomplishments that we're proud of
We built a working single-agent architecture where a natural-language family goal becomes a structured, persistent mission.
The agent can use existing family context instead of repeatedly asking the user for the same information, and clarification is determined dynamically rather than through a hardcoded questionnaire.
We also successfully deployed the backend on Google Cloud Run and persist family and mission state using Firestore.
What we learned
Building an agentic application is very different from adding an LLM to a chat interface.
Useful agents need reasoning, tools, persistent state, lifecycle management, and clear boundaries between probabilistic AI decisions and deterministic application behavior.
We also learned that autonomy does not mean giving the model control over everything. The strongest design gives Gemini responsibility for semantic reasoning while application code provides reliable execution boundaries.
What's next for Fam Agent
The next stage is connecting location and weather capabilities, nearby-store discovery, and Plan Break Recovery.
This will allow a persistent mission to react to real-world changes, such as weather or store availability, without requiring the family to restart the planning process.
We also plan to expand the shopping companion, product guidance, and household memory capabilities.
Our goal is for Fam Agent to move beyond answering family questions and become an agent that can stay with a household goal from planning through completion.
Log in or sign up for Devpost to join the conversation.