Agent infrastructureRecurlySoftware Engineer Intern2026Shipped
Recurly Agent Platform
Takes a product requirement and turns it into merged pull requests. Five agents do the work. A human signs off five times.
Problem
A written requirement becomes a merged pull request through a chain of handoffs: someone reads the document, someone splits it into work, someone writes the code, someone answers the review.
Most of the elapsed time is not any of those steps. It is the waiting between them. One agent handed the whole chain does not fix it either. It returns a diff large enough that nobody wants to review it.
Approach
I spent the summer at Recurly building internal agent infrastructure for the engineering and go-to-market teams, reporting into the platform group.
The through-line across all three things I shipped is the same: put an agent on the part of the work that is a handoff, and leave the judgement with a person.
How it works
Three things shipped between May and August, none of them customer-facing.
A product development lifecycle platform. A planner, a decomposer, a Sonnet and Opus coder pair, and a pull request responder, each an isolated Kubernetes Job through the Agent SDK.
An SDR assistant in Slack: async Python with three-tier Claude routing over BigQuery, 6sense, ZoomInfo and Salesforce, returning source-cited account briefs and outreach drafts.
A prospect report card pipeline, n8n to FastAPI on Cloud Run to Claude Opus, that turns an inbound form fill into a personalized PDF analysis of that prospect’s subscription business and emails it back.
Impact
The time came out of the waits between handoffs, not out of the review, so all five checkpoints stayed.
The Slack assistant answers with a source-cited brief in about twenty-five seconds.
Built with
Claude Agent SDK · Python · Kubernetes · FastAPI · BigQuery · Cloud Run
Built inside a company codebase, so there’s no public source to link. Happy to talk through the design.