Building AegisOps Autopilot with Qwen Cloud
Production incident response is a high-stakes workflow where speed and caution have to coexist. A thin chatbot is not enough. The system needs memory, tool-backed evidence, approval gates, and an audit trail that a human can trust.
AegisOps Autopilot is my Qwen Cloud Hackathon project for Track 4: Autopilot Agent. It accepts ambiguous production alerts, recalls prior operational lessons, gathers evidence from logs, metrics, change graphs, and policy checks, asks Qwen Cloud to produce a conservative diagnosis, and proposes a reversible remediation plan. Risky actions pause behind a human approval checkpoint instead of mutating production blindly.
Why Qwen Cloud Fits This Workflow
Qwen Cloud is the reasoning layer, not the entire system. AegisOps gives Qwen a bounded set of incident tools: log search, metric probing, change graph lookup, policy checking, and remediation simulation. When Qwen returns tool calls, the backend executes only incident-scoped tools, appends the results as tool evidence, and asks Qwen for the final diagnosis.
The live Alibaba ECS deployment now runs in Qwen Cloud mode through the DashScope OpenAI-compatible API, while the repository still keeps deterministic fixtures for repeatable judging and CI. That combination lets judges inspect both the cloud path and the reproducible fallback path.
What I Built
- A TypeScript Node API and React dashboard for incident triage.
- A Qwen Function Calling loop with five incident-scoped tools.
- Persistent operational memory and post-incident learning.
- An agent council with reliability, security, and human approval roles.
- Risk-scored remediation plans with reversible commands and verification metrics.
- OpenAPI and MCP surfaces backed by the same tool registry.
- Evaluation, ablation, model-ops, architecture, and deployment proof reports.
The Hard Part
The hardest part was not wiring up a model call. The hard part was keeping autonomy useful without making it unsafe. A production incident agent needs to say what it knows, where the evidence came from, what it is not allowed to do, and when a human must approve a change. That is why AegisOps treats evidence boundaries and reversibility as first-class product features.
The core lesson: an Autopilot Agent should not hide behind a confident answer. It should expose memory, tools, policy checks, approval state, rollback steps, and verification metrics.
Current Public Proof
The submitted project is live on Alibaba Cloud ECS at
http://101.201.33.56/. The proof endpoint at
/api/alibaba/proof returns Alibaba ECS runtime
metadata and Qwen/DashScope endpoint configuration without exposing secrets. The public workflow endpoint
has also been smoke-tested in qwen-cloud mode.
What I Would Improve Next
The next step is to connect AegisOps to real observability systems, expand incident classes, add stricter cost controls, and build team workflows for incident commanders and service owners. The product direction is clear: make high-risk automation inspectable enough for real operations teams to trust.