The infrastructure playbook
Everyone got Claude Code. Nobody had infrastructure.
Here's the honest version of what happened on my team. We gave everyone access to AI coding tools, and the creativity was incredible — people were building agents for everything. But in Friday reviews, what we'd see was localhost URLs. Brilliant work that lived on one person's laptop and died when they closed it.
Then it got riskier. Teams started buying their own Vercel accounts to host what they'd built — with no governance, no identity, no controls. Some were sharing information externally without realizing it. The building was working. The infrastructure wasn't there to make it safe, or to make it scale.
So I stopped and built the foundation first. This is what we put in place — and what I've learned marketing orgs most want to understand right now.
The journey
Localhost on Fridays
Where it started — Everyone had AI coding tools and was vibe-coding real agents. But the output was localhost URLs in our weekly reviews — great ideas trapped on individual laptops, impossible to share, use, or scale.
Shadow hosting, no governance
Where it got risky — Teams started buying their own Vercel accounts to actually host things. No authentication, no permission controls, no oversight — and in some cases, information was being shared externally. Speed without guardrails.
Infrastructure first
The turn — I decided to put the right foundation in place so the team could scale safely: a governed place to host, a place to discover each other's work, and a place to share code — so we stopped rebuilding the same things in isolation.
The foundation we built
Governed hosting
A central Heroku account hosting every agent experience behind VPN-enabled access — so nothing sensitive lived on a personal localhost or an ungoverned account again.
A PMM marketplace
An internal marketplace where my team could share projects — in production or still in ideation — so teams could see what already existed and stop building the same thing twice.
A central GitHub org
One shared repo for the whole marketing org. Teams could fork a project to extend it instead of reinventing the wheel — shared ownership, compounding work.
The non-negotiables
As we built the internal agent ecosystem, we established a few non-negotiables. This is the checklist I'd give any marketing leader standing this up:
- Feedback loops in every agent — Users can submit enhancements directly to the agent's owner — so every experience keeps improving.
- Analytics dashboards — Understand adoption, engagement, and real business impact — not just that something shipped.
- Authentication & identity — Know exactly who is using what.
- Permission controls — The right people have access to the right data and insights — and only those people.
- Secure hosting environments — Governed hosting instead of personal localhost deployments.
- Automated error logging — Errors are captured and reported automatically, not discovered by accident.
- Kill switches — Every experience can be shut off instantly if something goes wrong.
- Shared code ownership — Multiple maintainers per project — no single point of failure when someone changes teams.
- An internal marketplace — Employees can discover agents, share what they've built, and collaborate on new projects.
Want the deeper details? Let's connect inside Relay.
There's a lot more to how we built this — the actual architectures, the budget, the vendor calls, and what broke along the way. It's the kind of thing that's better as a conversation than a blog post. Join Relay, connect with me, and I'll share the full write-up on how we built the infrastructure.