AutoPM
multi-agent product management copilot
Project metadata
Overview
An AI product-management copilot that runs an end-to-end PM workflow inside a project "Space" using eight specialized LangGraph agents, producing real artifacts: user stories, RICE-prioritized backlogs, wireframes, stakeholder emails, and live Jira tickets.
Problem
Product managers juggle ideation, user stories, prioritization, OKR alignment, wireframes, emails, and Jira under time pressure. AutoPM offloads the mechanical portions to specialized agents.
The agents
All eight agents run on the Gemini API.
Additional surfaces: a context-aware AI assistant with access to project history, and a market research panel showing competitors, activity signals, and trends. The research panel deliberately avoids black-box sentiment scores and extracts themes and pain points instead.
Architecture
Every LangGraph node emits Zod-validated structured output, so one agent's result is directly consumable by the next without a translation layer.
Engineering challenges
- Cross-agent state: keeping each agent's output immediately usable by the next.
- Typed graph design: enforcing schema guarantees across LangGraph nodes.
- PDF chunking quality: reliable OKR extraction for the alignment step.
- API ergonomics: OAuth flows, rate limits, retries, and batching.
Full stack
LangGraph · LangChain · Gemini 2.0 Flash · Zod · Next.js 14 · React · TypeScript · Tailwind · Node.js · MongoDB · Mongoose · Auth0 · Tavily · Gmail API · Jira Cloud
Results
Placement | 2nd place, HackUTD 2025 PNC track | |
Participants | 1200+ | |
Agents | 8 named and described | |
Zod-validated typed state across all nodes | Yes | |
Live Jira ticket creation and Gmail sending | Yes |