Autopilot-agent

• 

  system workflow

Inspiration

Software teams ship fast, but manual release verification is slow, fragmented, and lacks clear proof. We built Autopilot-agent—an evidence-first AI agent that automates release checks with human approval at risky points.

What it does

• Plans & Executes: Converts PR diffs and natural language into structured Playwright browser/API test plans.
• Collects Audit Proof: Captures screenshots, network logs, and traces linked to immutable commits.
• Human-in-the-Loop: Enforces policy gates before running risky actions (deployments, PR updates).
• Release Recommendation: Delivers a clear pass/fail risk report for release managers.

How we built it

• Core App: Next.js 16 (React 19), Tailwind CSS.
• AI & Automation: Qwen LLM Gateway, Playwright in isolated workers.
• Durable State: Custom checkpoint system (FileCheckpointStore / Alibaba OSS layout) for resumable tasks.
• Cloud Infrastructure: Alibaba Cloud FC3 (Function Compute) with OSS & SLS logging.

Challenges we ran into

• AI Safety & Policy: Decoupling strict permission rules from prompt context so the AI can't bypass human approval.
• Long-Running Serverless Runs: Designing durable checkpoints so tasks survive function timeouts and process restarts.

Accomplishments that we're proud of

• Built a truly resumable agent that pauses for human approval and resumes without losing progress.
• Created immutable, audit-ready evidence packs tied to commit SHAs.
• Successfully deployed and verified end-to-end reference shop journeys on Alibaba Cloud FC3.

What we learned

• Prompts are not policy: Security guardrails must be handled by hard-coded policy engines outside the LLM.
• Evidence over answers: Teams need verifiable screenshots and logs, not just an AI saying "looks good."

What's next for Autopilot-agent

• One-click GitHub App integration with automated PR status checks.
• Multi-agent expansion (Security Scanner, Visual Regression Agent).
• Cloud adapters for AWS (S3/Lambda) and GCP (Cloud Run/GCS).

Built With

• alibaba
• langchain
• qwen
• typescript