Handoff Gate

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for MD DAYS

Inspiration

Multi-agent systems are powerful, but one agent’s mistake can silently propagate through an entire workflow. We wanted to create a trustworthy coordination layer where agents can collaborate without automatically trusting one another.

What it does

Handoff Gate coordinates a Researcher, Summarizer, and Formatter through validated artifacts. Every handoff is checked for structure, citations, provenance, and exposed credentials. Invalid work is blocked and returned for correction before it can reach the next agent.

Users can upload documents—including text-based PDFs—ask a research question, observe the live pipeline, and open the final cited Markdown report.

How we built it

We used React and TypeScript for the interface, Fastify for the control plane, Zod for artifact validation, and Codex CLI with Volcengine Ark for agent execution.

Agents operate in isolated workspaces. A coordinator brokers files between them, validates every artifact, records events and hashes, and manages rejection, retry, recovery, and stopping.

Challenges we ran into

The hardest challenge was enforcing provenance across multiple stages without relying on prompts alone. We also had to coordinate isolated workspaces, recover safely from rejected artifacts, preserve live session state, and extract usable text from uploaded PDFs without adding a large dependency.

Accomplishments that we're proud of

We built a working end-to-end pipeline where unsupported citations cannot reach the final report. The demo visibly rejects a fabricated citation, explains the violation, retries the stage, and completes with verified sources.

We also delivered live event tracking, persistent sessions, artifact hashes, document and PDF research, an enlarged report reader, and deterministic demo modes for reliable judging.

What we learned

Reliable multi-agent systems need more than capable models. They need explicit contracts, controlled information flow, observable state, and deterministic recovery.

Prompts guide agents, but validation gates provide guarantees.