Enabling unbounded SWE productivity are coding agents, popular ones like Claude Code, Kilo Code, OpenHands among many others. So how do you decide whether to use one of these at your company or build something of your own? Stripe’s blogpost on their new coding agent gives some reasons why you would want to customize a coding agent. In short, they have developed their own custom tooling from the ground up that their software engineers use on a day-to-day basis, battle tested. Their codebase is mostly written in Ruby which requires different tooling. And when your code grows to such a big volume, you require special handling of context given the limited window that you have while operating an LLM.
What’s interesting is that this is not really a from-scratch development of an agent. They forked an early version of Goose Agent by Block and then built on top of that. What they are mostly building is harnesses for the agent. The agent framework, you could pick up one of the popular open source ones out there, but it is the harness and your company-specific context that makes the agent useful. So no, you do not need a custom agent implementation, but what you do need is thinking about harnesses for that agent, skills for that agent.
What’s interesting about Stripe’s implementation of their coding agent is they have also created an MCP server with over 400 tools to their internal SaaS platform and other internal documentation tools. So the agent has a powerful tool belt that it could use to gather context from all across the company. What stands out to me is that building these tools might be the core part of Minions and how they work, because the agent loop is sort of set but it is the tools that dictate the information to the agent itself.
Harnesses also provide key feedback to an agent, and that’s where agent creativity comes in.