Inspiration
Real business teams drown in repetitive, multi-step workflows: a customer email arrives, someone reads it, drafts a quote, checks it, and sends a reply. We wanted an Autopilot Agent that turns a plain-language intent into a structured, runnable workflow — while keeping a human in the loop at the decisions that actually matter. Qwen Cloud's strong instruction-following and tool-use made it the natural reasoning engine for this.
What it does
Automatom lets you describe what you want in plain language and scaffolds an executable workflow from it. It exposes a minimal REST surface (POST /workflows, POST /runs, GET /runs/:id) backed by a persistent store and a pluggable step engine. Steps can be:
- LLM steps — powered by Qwen models on Qwen Cloud for reasoning, drafting, classification, and extraction
- HTTP steps — call external tools and business systems
- Code steps — deterministic transforms
- Approval steps — human-in-the-loop checkpoints that pause a run until a person approves, matching the Autopilot Agent track's emphasis on production-readiness over toy demos
Every run is persisted with status, timestamps, and per-step output so workflows are auditable.
How we built it
The backend is a FastAPI + Pydantic service. The workflow engine executes steps sequentially, persisting each run. The LLM step calls the Qwen Cloud API (DashScope-compatible OpenAI endpoint) so any Qwen model can drive reasoning. The backend is containerized and runs on Alibaba Cloud.
Challenges
Designing the approval-step semantics so a run can pause and safely resume, and keeping the agent's tool-calling reliable on ambiguous inputs, were the hardest parts.
What's next
Parallel step execution, richer tool connectors, and a visual workflow builder.
Log in or sign up for Devpost to join the conversation.