FlowRouter Safe Agent Gate

A runtime gate that routes AI-agent actions into ALLOW, VERIFY, STOP, or REROUTE before execution.

Comment

Inspiration

Most agent demos focus on what an AI can do.

FlowRouter Safe Agent Gate focuses on the missing step before execution: should the proposed action still be allowed?

AI agents can reason, plan, and use tools. But planning is not permission. Before an action becomes real, the system should be able to allow, verify, stop, or reroute it.

What it does

FlowRouter Safe Agent Gate routes proposed AI-agent actions through a runtime decision layer before execution.

The gate classifies each proposed action as:

  • ALLOW
  • VERIFY
  • STOP
  • REROUTE

The demo focuses on GitLab workflows. A Gemini-powered agent proposes a GitLab action, such as creating an issue, drafting a comment, or preparing a follow-up task.

Before the action can execute, the gate evaluates context, risk, and execution intent.

How we built it

The project is built as a small public proof-of-concept.

The initial implementation includes:

  • A FastAPI backend
  • Agent proposal and gate evaluation endpoints
  • Mock execution
  • An in-memory audit trail
  • Tests for gate decisions
  • Clear adapter boundaries for Gemini, Google Cloud Agent Builder, and GitLab MCP

Challenges we ran into

The main challenge is separating helpful agent planning from actual execution authority.

A confident agent response is not the same as permission to act.

Accomplishments that we're proud of

We built a working public proof-of-concept that demonstrates the core control primitive: agent proposes, gate decides, execution waits.

The backend includes deterministic gate behavior, mock execution safeguards, and tests for ALLOW, VERIFY, STOP, and REROUTE.

What we learned

Agent safety is not only about better answers. It is also about better boundaries before actions happen.

Planning is not permission.

What's next for FlowRouter Safe Agent Gate

Next steps are to connect the proposal layer to Gemini and Google Cloud Agent Builder, integrate GitLab MCP for real project context, host the demo, and show a clean decision trace for every proposed action.

Built With

Share this project:

Updates