AI Skills Research & Strategy: Training Coding Agents to Use MongoDB Best Practices

Role

Product Design Lead

Company

MongoDB

Timeline

Q2 2026

Overview

Goals

Support developers by improving AI recommendations for developing with MongoDB. Evaluate the MongoDB AI Agent Skills to understand usage pain points and content gaps. Better understand the user journey of a developer using coding assistants to work with their database in order to prioritize the highest value work.

Core Team

Product Manager, Software Engineers, and me

8

Interviews

1

Hour Each

4

Reprioritized projects

1

New team created

Problem

Business Context

MongoDB's flexible document model is a strength for developer velocity, but can be a risk if setup is done hastily or incorrectly. Many teams put schema design decisions directly in developers' hands, and developers optimize for short-term agility over long-term structure. Teams that use schema design "anti patterns" convert to paying customers 3.4x faster than teams that don't, but have significantly lower long-term retention. This ultimately leads to performance issues for our customers and reputational and financial risk for MongoDB.

As AI coding agents (Github Copilot, Claude Code, etc.) increasingly make and influence those structural decisions so it has become increasingly important to provide guidance to these agents to ensure they're using MongoDB best practices.

We needed to move quickly in order to fill the expertise gap for users so we shipped a set of agent skills with minimal up front research. We needed to better understand what users expected from our skills, what pain points they were experiencing, and what next steps we could take to best support their work.

Constraints

We don't control the agent's response: The nature of AI agents is that their responses are deterministic, thus we cannot fully account for what a coding agent might suggest to the user. We can only provide additional context to improve upon the agent's default response.

Rapidly changing technology: Given the speed at which developer processes are changing in the age of AI we needed to move quickly. We also understand that the landscape is continuously shifting and we need to frequently revisit any findings from this research.

This (skill) spotted issues that I was not even aware of.

Developer with 10+ years of experience

The image featured at the top of the about us page #2

Research

User Demos

Participants shared their screens and used the actual MongoDB MCP server and schema recommendation skill against their own schemas or a test cluster.

8 MongoDB users (developers, CTOs, engineering managers)
60 minutes
Live demos using the MongoDB MCP server
Using Claude Code, VS Code with Copilot, or Cursor

Customer Conversations

I also drew on direct enterprise customer conversations (joining calls with our customer success team) to validate that the friction points I was seeing in the demos matched what customers were experiencing in their work.

finding 1

Users are not aware of MongoDB skills

None of the participants had heard of our agent skills prior to the demo

finding 2

Users struggled to trigger the skill

5 of 8 users were unsuccessful at triggering the skill with their first prompt

finding 3

Security settings are not one size fits all

Users have vastly different levels of comfortability with AI having access to their data

finding 4

Users are still defining their AI processes

Engineering leads mentioned their teams don't have any one way of using these agents and processes are still being defined

Roadmap and product Solutions

Meet developers where they are by continuing to invest primarily in third-party AI tooling

Developers are increasingly making database decisions inside tools like Claude Code, Cursor, and VS Code, not inside MongoDB's own GUI. One participant put it plainly: "I consider it a good week/month/year if I don't have to log into MongoDB Atlas."

Guide users to next steps, don't invest in agentic solutions (yet)

Users were explicit that they don't want AI touching production databases directly; they see the skill as supporting planning, not execution.

We're working on improvements for the skill to surface: a to-do list, migration code, estimate ROI, and provide additional rationale.

Prioritize improvements for skill discovery and triggering

5 of 8 participants failed to trigger the skill on their first prompt and none were aware of our skills prior to our interview. No amount of improved advice matters if users can't reliably reach the skill in the first place.

We used language directly from our user interviews to expand how skils were triggered and are working on higher skills visibility on several surfaces for our plug-ins and MCP server.

Allow users to choose what they share with agents

Users had sharply divided comfort levels: some were fine with MongoDB automatically pulling read-only workload data, "we're already trusting you with a lot of things", while others (especially in highly regulated industries) had a hard "no" to any automated telemetry access.

We chose to pursue a skill design that accommodates both personas with follow up prompts.

Impact

Discovering What Not to Do

While we gained valuable insight into how users are interacting with MongoDB agent skills, this research was perhaps the most impactful by informing the team what we should NOT focus on. This research helped us to pivot from several high effort projects within our first party tools. Instead we're continuing to invest more resources in a newly created AI Builder team.

AI Supported Research

I partnered with our research operations team to test a few Beta features for AI run interviews. This would help us to become much more efficient as a team by generating insights without a dedicated interviewer. We ran into several issues with the features, especially as it related to our technical concepts. I will continue testing and providing feedback.

reflection

Next Steps

We need to continue to invest in AI supported research so we can move more quickly and validate our team priorities as the AI landscape continues to disrupt developer workflows. I will continue to partner with our research tools' customer success teams and provide feedback on their Beta solutions.

I will also focus on MCP applications to bring helpful UI patterns into 3rd party tools. This will support users in turning our recommendations into actionable next steps by providing previews, to-do lists, and clear ROI for schema design changes.