Ramp, the corporate card and finance software company, now runs every one of its AI agent products on a platform it built itself, and AI agents wrote the first version of that platform’s code. Sam Kronick, the engineer who led the work, described the project on Ramp’s engineering blog on 20 October 2026. The platform is called Laminar, and its first version took six weeks.
Start with what Laminar does. An AI agent is a model that uses tools over and over until a job is finished. The “agent harness” is the software wrapped around that cycle: it remembers the conversation, hands the model its tools, keeps the job running if a server restarts, and records everything for later review. Kronick says the simplest harness is a bare loop, but a production one grows to cover sandboxes where agents run code, subagents that take on pieces of work, and live streaming to users.
Ramp’s agents are varied. Kronick lists a support chat assistant called “Ask Ramp”, an invoice-reading agent inside Bill Pay, a coding agent named Ramp Inspect, and Glass, a shared agent coworker that can work for minutes or hours and report back in Slack. Each type puts different demands on the platform, which is why Ramp wanted one foundation instead of a dozen.
The choice to build came after a long stretch of buying and borrowing. According to Ramp’s post, the team tried model makers’ agent kits, managed agent services, open source tools, popular frameworks, and several homegrown experiments. Around May 2026 it decided to consolidate. Kronick gives three reasons for not simply taking a vendor’s product. Ramp wanted to switch between AI models freely, and a single provider’s service would lock it in. It wanted the platform team to sit beside the product teams and ship in minutes. And it wanted to own the debugging tools, because Kronick says good agents come from reading their step-by-step logs, something Ramp’s chief technology officer is fond of printing out on paper to spot waste.
The fourth reason is the one that applies beyond Ramp. Kronick writes that coding agents had improved enough that owning and upkeeping a platform cost little compared with taking on vendor dependencies. That is a company’s own accounting, and the post gives no price tag or headcount to check it against. Still, it describes a real shift: in-house software that once needed a large team can now be started by a handful of people.
Kronick is careful about how the code got made. He rejects the label “vibe coded.” A small team of experienced engineers feeds the agents design ideas, challenges the output, and reviews every change from colleagues against the overall architecture. For its foundations Ramp picked dull, proven tools (Postgres, Temporal, and S3) because its engineers already know how to run them and repair them when they break.
The closing lines carry the real caution. Kronick says he would do it again, and that a small team goes surprisingly far with today’s coding agents. He adds that doing this in-house has never cost less, yet people are still required: engineers who carry the pager, keep the system healthy, and learn what the product teams truly want from it. Ramp’s account comes from its own blog, with no outside audit of uptime or cost.
For any company weighing build against buy in agents, the arithmetic of the first draft has changed, and the bill that remains is for people. The thing to budget for is the on-call rota, not the six weeks.
Reported by Ramp on 20 October 2026, in its engineering blog post “Building Laminar: Ramp’s Managed Agents Platform” by Sam Kronick.
