AgentOps Commander

AgentOps Commander is a Gemini and Google Cloud Agent Builder-ready, human-supervised incident response agent demo for the Google Cloud Rapid Agent Hackathon.

Repository: https://github.com/RayWu1037/agentops-commander

Hosted demo: https://raywu1037.github.io/agentops-commander/

Recording console: https://raywu1037.github.io/agentops-commander/recording.html

Submission release: https://github.com/RayWu1037/agentops-commander/releases/tag/v0.1.0-submission

Devpost submission: https://devpost.com/software/agentops-commander

Final Submission Status

Item	Status	Evidence
Public repository	Ready	MIT licensed GitHub repo
Submission release	Ready	v0.1.0-submission
Hosted demo	Ready	GitHub Pages with post-deploy smoke tests
Partner track	Ready	Arize / Phoenix MCP
Gemini + Agent Builder contract	Ready	agent-builder-config.json
Final external tracker	Open	GitHub issue #1: #1
Demo video	Ready	https://youtu.be/TrcCUcYIb6c
Devpost final submit	Submitted	https://devpost.com/software/agentops-commander

Judge Quickstart

1. Open the hosted demo.
2. Click Run agent.
3. Watch the agent plan, call tools, identify root cause, and stop at the approval gate.
4. Approve the action.
5. Open Trace, Evals, and Architecture.

For a requirement-to-evidence review path, see JUDGES_BRIEF.md and SUBMISSION_AUDIT.md.

The demo shows the loop the submission will emphasize:

1. Receive a real-world operations incident.
2. Plan the investigation.
3. Call structured tools.
4. Explain evidence and root cause.
5. Require human approval before state-changing actions.
6. Emit Phoenix-style traces and evaluations.
7. Use the trace review to improve the next run.

Why this can win

Most hackathon agents stop at tool calling. AgentOps Commander is framed as a production agent: it takes action, but it is observable, evaluated, and gated by a human approval step. That maps directly to the judging criteria for technical implementation, design, impact, and creativity.

Run locally

npm start

Open http://localhost:8080.

On Windows PowerShell, if npm.ps1 is blocked by execution policy, run:

node server.js

You can also open the static demo directly:

public/index.html

No API keys are required for the demo mode. The backend returns deterministic agent runs so the demo is reliable for recording.

Quality gate

node scripts/test.mjs

The current suite covers security headers, path traversal payloads, HTTP verb tampering, malformed requests, XSS-shaped payload escaping, offline fallback, accessibility metadata, mobile layout risks, submission assets, and the agent approval flow.

GitHub Actions runs the same checks on push and pull requests. A Pages workflow deploys the static demo from public/.

Offline demo mode

The front end can run directly from public/index.html. If the API is unavailable, the app automatically falls back to deterministic client-side data, so the demo flow still works for recording and review.

Submission assets

• project-gallery.png is a 3:2 image suitable for the Devpost image gallery.
• public/recording.html is a hosted video recording console.
• public/demo-captions.srt is the hosted caption file linked from the recording console.
• DEMO_VIDEO_ASSEMBLY.md contains the final Seedance-style generation and editing plan.
• SEEDANCE_DEMO_PROMPTS.md and seedance-demo-prompt-pack.txt contain optional AI video prompts for intro/outro clips.
• JUDGES_BRIEF.md contains the fastest judging path through the demo and repository.
• JUDGE_QA.md contains concise answers to likely judge objections.
• DEMO_SCRIPT.md contains the three-minute video narration.
• demo-captions.srt contains upload-ready English captions for the demo video.
• VIDEO_QA_CHECKLIST.md contains the pre-upload video recording checks.
• npm run record:demo can produce a browser-recorded .webm walkthrough when Playwright and Chrome/Edge are available.
• DEVPOST_SUBMISSION.md contains copy-ready Devpost text.
• FINAL_SUBMISSION_PACK.md contains copy-ready Devpost fields, links, and video compliance checks.
• SUBMISSION_AUDIT.md maps the project to requirements and judging criteria.
• agent-builder-config.json defines the Gemini, Google Cloud Agent Builder, Arize Phoenix MCP, safety, and eval contract.
• TESTING.md contains the regression checklist.
• .github/workflows/ci.yml runs the quality gate.
• .github/workflows/pages.yml deploys the static demo to GitHub Pages.

Google Cloud and Gemini integration plan

The local runtime is intentionally code-owned so it can be moved to Google Cloud Run and wired to Gemini, Google Cloud Agent Builder, and Arize Phoenix MCP without rebuilding the product surface. The concrete agent contract is captured in agent-builder-config.json.

Planned production components:

• Gemini / Google ADK for reasoning and tool selection
• Cloud Run for the web app and agent runtime
• Secret Manager for API keys
• Firestore for run state and approvals
• Arize Phoenix for traces, spans, evaluations, and datasets
• Phoenix MCP for agent self-review over previous traces
• OpenInference / OpenTelemetry for instrumenting model and tool calls

Environment variables

GOOGLE_API_KEY=
GOOGLE_CLOUD_PROJECT=
PHOENIX_COLLECTOR_ENDPOINT=
PHOENIX_API_KEY=
PHOENIX_MCP_SERVER_URL=

Deploy to Cloud Run

gcloud run deploy agentops-commander --source . --region us-central1 --allow-unauthenticated

The repository includes a Dockerfile for container builds, but Cloud Run source deploy also works for this Node runtime.

Demo video outline

1. Show the incident page and metrics spike.
2. Run the agent.
3. Walk through the plan and tool calls.
4. Show the root cause and approval gate.
5. Approve the mitigation.
6. Switch to traces and evals.
7. Explain the Phoenix MCP self-review loop.

Keep the final video under three minutes.