Inspiration
Professionals repeatedly lose time to the work between a request and a decision: preparing weekly status updates, synthesizing changes, and producing a first recommendation. A chat response is insufficient because the user needs work to continue in the background and remain reviewable before anything consequential happens.
What it does
BriefRunner accepts a plain-language briefing request and starts an asynchronous, inspectable workflow. In its Google-native production path, the agent uses Gemini 3.5 Flash through Vertex AI to draft a concise brief with the audience, findings, uncertainty, recommended next action, and an explicit approval checkpoint. The API exposes run state for review. Approval records an approved handoff only; it never sends a notification automatically.
How we built it
The FastAPI backend was deployed on Google Cloud Run. BriefRunner invokes Gemini 3.5 Flash through Vertex AI using the Google Gen AI SDK, the Google agent framework used by this project. A small SQLite store persists the demonstration run state and status endpoints expose its progress. The code deliberately does not claim live data without being given it, execute arbitrary tools, or make irreversible changes.
The repository includes a Dockerfile, dependency manifest, architecture diagram, local spin-up guide, Cloud Run deployment instructions, tests, and a non-secret health endpoint. The demo video distinguishes a reproducible offline review run from recorded Cloud Run evidence: it shows the deployed service URL and historical successful Gemini-mode request logs, then the returned brief, approval transition, and the fact that sent remains false.
Challenges we ran into
The core challenge was preserving credible asynchronous workflow behavior while keeping the action boundary honest. The implementation separates the request, background work, persisted status, reviewable output, and approval transition. We also migrated the runtime to the supported Google Gen AI SDK Vertex AI configuration and removed the legacy AWS/Strands path from the submission repository.
Accomplishments that we're proud of
We built an inspectable agent workflow rather than a generic chat screen. BriefRunner makes state and uncertainty visible, returns control to a person before external action, and gives reviewers a reproducible path to verify the Gemini, Google Gen AI SDK, and Cloud Run integration.
What we learned
An agentic workflow is most useful when it makes progress without hiding what happened. By making approval and sent: false observable in the API result, the system demonstrates background work while preserving a human decision boundary.
What's next
The next production step is replacing the demonstration’s local run-state store with managed persistence and adding narrowly scoped authenticated connectors that act only after approval.
Pre-existing work disclosure
BriefRunner’s Google-native agent workflow, Gemini-on-Vertex-AI adapter, Cloud Run deployment, architecture, tests, and submission materials were built during the Submission Period. I incorporated my own pre-existing generic Automatom workflow scaffolding—basic FastAPI schemas, local run-state persistence, and non-agent workflow routes—as a disclosed starting point. The reused code is owned by me and MIT licensed; the Google-native BriefRunner functionality submitted here is the new work. The repository contains PRE_EXISTING_WORK.md with details.
Log in or sign up for Devpost to join the conversation.