I’ve been building tools for AI agents at Arcade.dev since before MCP existed, and we make a lot of them (8,000 and counting). But it seemed like no matter how efficiently we worked, it wasn’t fast enough. So I ran an experiment. Could I hand agents years of knowledge, speak what I wanted, never type or read a line of code, and have them build a system to efficiently and effectively build MCP servers? The answer was yes. We call it Gauntlet.
The agent is the consumer of the tools
The insight is simple once you say it. The agent is the one that uses the tools, so the agent should be the one that tests them.
Gauntlet frames an agent as a power user of the underlying service and turns it loose. The agent is told to do something a real user would actually want, and to report back bugs and oddities that surfaced during that exploration. It gets no manual for the tools, so it has to figure them out the way a real user would.
Give it a Google Slides toolkit, for example, and it goes off building slides about whatever it likes. As it works, it surfaces the rough edges that only appear in real use. A second agent rewrites the code from those findings, while a third re-runs a locked set of scenarios to ensure there isn’t any regression. Then the loop runs again.
This runs for about 10 hours, so I can set it up before bed and wake up to roughly 10,000 lines of code that has reached convergence, meaning the loop stops finding rough edges, and a different model reviewing the final diff signs off too. What began as a weekend project now runs in our daily work.
For example, over a recent weekend Gauntlet produced 10 new toolkits, work that used to take our team about two months. We rewrote the Twitter toolkit and built new ones for Google Slides, Fireflies, Insightly, and Datadog, among others.
Why this matters for customers
A customer recently needed an MCP server for Cursor agents. We had it running in their production environment the next day. That turnaround, from request to working tool, is something we couldn’t offer before.
It also changes how we think about coverage. We don’t want an endless catalog of services no one uses. We want the connections customers depend on, like Google Workspace and Microsoft 365, to stay current, optimized, and bug-free. Gauntlet lets us keep raising that bar instead of chasing breadth.
What’s next
Gauntlet still has rough edges. We’re still tightening the review loop, and a human reviews a large volume of generated code before anything ships. The newest piece is provisioning: automating the creation of OAuth apps that every tool talking to an upstream service needs to reach production. That works for a couple of providers today, but the current work is making it reliable enough that the whole team can depend on it.
But Gauntlet already shows what becomes possible when we point agentic AI at the work of building agentic AI. It is how we keep the tools our customers depend on current and reliable, and how we get new capabilities into their hands faster than we ever could before.
