Inspiration
The software development lifecycle is undergoing a fundamental shift.
With the rise of AI coding agents, developers are no longer writing every line of code rather they are guiding intelligent systems to generate it. However, this shift has exposed a critical gap:
AI agents are powerful, but they lack project-specific intelligence and contextual awareness.
In real-world teams:
- Coding standards are implicit and scattered
- Design decisions are distributed across tools like Jira and Confluence
- Critical insights emerge during code reviews and team discussions
In today’s AI-driven workflow:
- Developers repeatedly re-teach the same project context and standards to their AI agents
- AI-based reviews often focus on syntax and correctness, missing deeper aspects like architecture, intent, and business alignment
As a result, knowledge remains fragmented, and AI systems fail to truly align with how teams build software.
This led to a key realization:
What if we could build a system where agentic skills evolve continuously, and every code review understands the full context behind the code?
DuoMind was born from this idea : an agentic brain that evolves with the entire development lifecycle, continuously building agentic skills from code changes and making reviews context-aware by incorporating design, requirements, and team discussions.
What it does
DuoMind ensures AI doesn’t just write code, but understands how your team builds and reviews software
It works by:
Building evolving agentic skills
- Learns coding patterns and standards from the codebase and merge requests
- Eliminates the need to repeatedly guide AI with project-specific rules
Making code reviews context-aware
- Uses Jira and Confluence to understand requirements and design
- Reviews code based on intent, not just syntax
- Uses Jira and Confluence to understand requirements and design
Capturing knowledge from discussions
- Extracts insights from Slack conversations
- Converts them into structured review comments
- Extracts insights from Slack conversations
How we built it
DuoMind is built on the GitLab Duo platform using agents and flows, combined with contextual integrations.
1. Agent Skills Builder [Gitlab Duo Agent]
- Scans the codebase to identify recurring patterns and conventions
- Extracts reusable coding practices
- Generates an evolving skills repository that can be used by AI coding agents (e.g., Anthropic's Claude)
2. Agent Skills Builder Flow [Gitlab Duo Flow]
- Continuously monitors Merge Requests for meaningful changes
- Detects patterns worth adding to the agentic skills
- Automatically updates the skills within the same MR, reducing manual effort for developers
3. MR Review (Jira + Confluence) [Gitlab Duo Agent]
Combines:
- Code changes
- Context from Jira tickets and Confluence documents using Atlassian's MCP
- Code changes
Generates:
- Context-aware review comments
- Feedback aligned with design and requirements
4. Slack Meeting Notes Synthesizer [Gitlab Duo Agent]
- Extracts relevant discussions from Slack channels related to the MR using Slack's MCP
- Processes meeting notes and conversations
- Converts them into structured review comments
Challenges we ran into
- Identifying useful patterns from MRs
- Many code changes are one-off or too context-specific
- Deciding what qualifies as a reusable “skill” was tricky
- Integrating skill updates into MR flow
- Updating skills within the same MR without disrupting developer workflow
- Ensuring the automation feels helpful, not intrusive
- Fetching and aligning external context
- Mapping Jira tickets and Confluence docs correctly to the right MR
- Handling missing or loosely linked context
- Linking Slack discussions to code changes
- Hard to reliably associate conversations with specific Merge Requests
- Extracting only relevant parts from noisy discussions
- Hard to reliably associate conversations with specific Merge Requests
Accomplishments that we're proud of
- Reimagined AI-driven software development
- Moving from prompt-based coding to systems that learn and adapt
- Moving from prompt-based coding to systems that learn and adapt
- Elevated code reviews beyond syntax
- Bringing context, intent, and design into AI-powered reviews
- Bringing context, intent, and design into AI-powered reviews
- Introduced continuous learning into the dev lifecycle
- Enabling systems that evolve with real development workflows
- Moved towards truly team-aware AI systems
- AI that understands how teams build, review, and improve software
What we learned
- Importance of agentic skills
- Realized that evolving agentic skills are key to making AI truly useful in real projects
- Realized that evolving agentic skills are key to making AI truly useful in real projects
- Value of external context via MCP
- Integrating Jira, Confluence, and Slack significantly improves the quality of AI outputs
- Deep understanding of GitLab Duo platform
- Learned how to build agents and flows to automate and enhance the development lifecycle
What's next for DuoMind
- Richer, unified context layer
- Expanding beyond current integrations to build a single source of truth across code, docs, and conversations
- Expanding beyond current integrations to build a single source of truth across code, docs, and conversations
- From project intelligence → organization intelligence
- Extend DuoMind to learn and enforce standards across repositories, enabling consistent engineering practices at scale
Built With
- duo
- gitlab
- mcp
Log in or sign up for Devpost to join the conversation.