Agentic development for my team

The problem

Every team at Contentful had access to Claude Code and a mandate to use it to speed up work. How it fitted into each team’s workflow was left to the team. My teammates were using it to write code, and I could see it doing a lot more with the right setup.

What I built

I’d already connected Claude to most of my personal projects: a knowledge store, Hugging Face, GitHub. At work I took stock of what we had. Jira and Confluence, incident tracking, logging and metrics, other agent systems. Each one opened up something new. The MCP connectors existed, but very few people used them, and fewer had thought about connecting them all together.

I wired them up through MCP and SSO, then wrote skills to orchestrate them. One skill picks up a Jira ticket by number, pulls the plans and specs from Confluence, checks what already exists in GitHub, and looks at metrics, incidents and logs before shaping a feature or a fix. Most of the value is in that context gathering. Work starts from what the team already knows.

People stay in charge

Agents gather context, plan and draft. Review, approval, merging and deployment stay in human hands.

Getting the team on board

Investigating, debugging and fitting the connectors and skills together took me a while. A markdown doc listing the steps wasn’t going to persuade anyone to try it. So I did what any good developer would do. I automated it, and shared a setup script you run through with an agent.

What changed

The team uses it daily. Teammates have added their own skills for tech health and standups, and it’s become the standard way my team does agentic development. It’s one of several AI initiatives at Contentful. The part I like most is that other people are now building on it.