I created this post for the purposes of entering the All Things Agentic hackathon on Devpost, a Google Cloud hackathon about building real agents. The deadline is August 31, the demo is already live, and the entry goes in before the deadline. This is the story of the sprint.
Amalthea is a meal planning agent that interviews the household before it plans anything. It asks which nights are busy and which meals are fixed, then builds the week from the pantry, the budget, and the allergies it already knows. Every grocery list is computed as plan minus pantry, so the family uses what it has and buys only the gap.
The whole entry was built solo, from scratch, inside the submission window, in its own public repo. Gemini 3.7 Flash through Vertex AI, the agent runtime on ADK for TypeScript, the service on Cloud Run, household memory in Firestore.
Two design calls carried the build. First, the agent has exactly eight tools and they are the only write path. The model cannot change anything by talking. Second, allergen exclusions live in a deterministic lookup table outside the model. The model suggests, the table blocks. A recipe carrying an allergen the table recognizes cannot enter the plan no matter how the model is asked. It is an assistive filter on the household's own list, not a guarantee.
The honest part of building in public: before recording the demo I ran an independent review against the live service, and it found real bugs. The nastiest one made repeated meal swaps bounce between two recipes forever, which would have failed on camera. Six defects got fixed before recording, and the ones that could show on camera are covered by tests. The repo carries 75 passing tests plus a written adversarial testing report, four attacks documented with their results.
The demo is live on Cloud Run and the repo is public:
- Try it: https://amalthea-demo-947902910401.us-east1.run.app
- Code: https://github.com/Morain-Labs/amalthea-agent
The judges score the video on unedited live execution of the agent, which is the kind of judging criterion that keeps a hackathon honest, and it is how the demo will be recorded.