How Coursera Became an AI-Native Enterprise

What’s an AI-native enterprise? Can a decades-old public company become one?

Coursera is a fourteen-year-old public company that has transformed itself to the leading edge of AI, and in doing so redefined how education and engineering can be done better with AI: personalized assessments and AI tutors for learners, a Course Builder for its three-sided marketplace, and a custom reinvention of the agentic engineering SDLC inside Mustafa’s team. Coursera is turning AI into a growth engine with courses on exactly the skills companies are looking for in new hires. In May 2026 it completed a $2.5 billion combination with Udemy, and the two platforms now reach more than 200 million learners.

I sat down with Mustafa Furniturewala, CTO at Coursera, and his MCP Gateway technology partner Jiquan Ngiam, CEO of MintMCP, to discuss how they did it. It comes down to three things: enablement, culture, and automation.

1. Enablement

Being AI-native means empowering the entire company, and the benefits of AI come with risk. Engineering is hard enough to transform when you have decades of technical debt. But being truly AI-native means getting everyone in the company to become an expert user, not at all easy to do securely in a large enterprise.

“The primary goal is to enable everyone. My first instinct is not to block access, but to see what is the secure, right way to enable people with the tools.”

Before MintMCP, API tokens were scattered across laptops, business users wanted Claude Code access, and nobody could say which MCP tools were approved. MintMCP is a gateway that governs AI’s access to data. It centralizes tokens so no keys sit on laptops, bundles tools by role, scans inline for sensitive data, and keeps audit logs. IT sets the policies once and they work across Claude, Cursor, and ChatGPT etc. with a unified view of every agent.

Enablement is also teaching AI how to work. A skill is a markdown file of standard operating procedures. Coursera forked Anthropic’s knowledge-work skills repo and built many skills. For example, an onboarding skill for new engineers replacing the Confluence pages that used to hold that process.

2. Culture

Coursera built an inclusive AI culture. There is an AI council, a group of engineers that meets to discuss the company’s AI frameworks and how to improve them. There is a plugins repo everyone contributes to, an award for the best plugin of the month, and Spark sessions where people show each other how they use AI, which sparks more plugins in turn. A developer experience team runs much of this.

More than a quarter of Coursera’s Claude usage now comes from people who do not write code. Cowork edits the document, the spreadsheet, and the deck directly and pulls the data itself. Those business users are contributing skills back. Mustafa put the mindset this way:

“I don’t think we are fully there… But the key is to have a system, keep improving it, and create feedback loops where the value compounds over time.”

3. Automation

Mustafa’s team is creating their new agentic SDLC. He says it is a great time to be an engineer, because AI takes the grunt work across more than 200 repos and millions of lines of tech debt built up over a decade, leaving design and product taste to humans.

A Scala-to-Java migration that took engineers weeks now runs in two to three hours, deployment included. A separate migration of 500 tests ran in weeks instead of quarters. It doesn’t just work by default though:

“If you just spin up Claude Code and just do this long-running thing, it’s likely going to fail, or it will hallucinate, it will create bugs. So the setup and the system around it becomes very critical.”

The setup around the agent for Coursera means nested CLAUDE.md files with the architecture and patterns written down, and a model that retrieves the right context itself rather than through a separate index.

At MintMCP, Jiquan goes further. A review loop has Claude write a plan and the code, Codex review it with fresh eyes, and the two go back and forth until the code is simpler and secure; sometimes the pull request gets smaller. This “Do-Anything” loop or Dan-loop takes a while, so JQ moved it into a sandbox triggered from Slack.

Jiquan runs a whole team of agents: one recreates his exec team to debate a document, and another, called Charlie, reads the week’s sales calls to tell the team what is and is not selling. The agents do real work across functions, each scoped and read-only where it should be.

The tech: API, CLI, or MCP

Jiquan explains why MCP, when APIs, CLIs, and browser use exist. APIs aren’t designed for agents and may not allow the level of granular access controls or have the usability needed. MCP moves that work to the server, so a business user never touches a command line. Notion’s first MCP copied its API one-to-one; the second was rebuilt for agents, with a new language, more explicit granular scopes, and works far better.

Ultimately agents are powerful only when they have the right tools. MCP Gateways are a secure way to enable agents for all sorts of users, not only engineers.

Clips from the episode: YouTube playlist

📖 Full episode with timestamps, key takeaways, and show notes.