Every employee at Kapwing - including the sales reps, content writers, and customer support staff - committed code to their production codebase in Q1 2026. Not toy projects. Real pull requests.
The video editing startup has about 25 people total, 12 of them engineers. CEO Julia Enthoven set the goal at the start of the year: 100% of the team using AI coding agents by the end of Q1. They hit it on March 23rd.
This is a real-world test of something a lot of companies are talking about but few have actually executed - using AI coding agents (tools like OpenAI's Codex that write, test, and submit code based on plain-language instructions) to collapse the wall between technical and non-technical roles.
The Setup
They ran primarily through Codex, OpenAI's coding agent, with Claude Code as a secondary tool. The infrastructure connected Codex to Slack and GitHub, plus their bug tracker Shortcut. The key unlock was one-line deployment: a non-technical employee could trigger a production deploy by leaving a comment on a GitHub pull request.
Rollout happened in three phases across five months. From October to November, engineering led infrastructure setup and gave QA and product managers first access. February was about expanding permissions and getting dev machines configured for non-technical staff. Training happened mid-February, led by the CEO, and covered how to prompt an agent, how to iterate on its output, and how to get something deployed.
That sequencing mattered. They didn't hand people a tool and wish them luck. They pre-assigned well-scoped stories - small, clearly defined tasks - to each employee before the training session, so everyone showed up with something concrete to work on.
What Non-Engineers Actually Shipped
The outcomes are where this case study gets specific. Kapwing's QA manager ranked in the top five employees by number of PRs submitted in Q1. The content team started building interactive HTML visualizations - things that previously would have gone into an engineering backlog. A sales rep resolved enterprise accessibility issues that had been deprioritized for months.
They logged nearly 108 pull requests tagged with Codex in Q1 alone. They also eliminated their quarterly "bug bash" events - sessions where engineers set aside normal work to clear a backlog of small bugs. Those events cost roughly 36 engineering days per quarter, which the CEO calculates as about half an engineering salary saved annually.
Production incidents decreased compared to previous quarters.
The Part That Surprised Them
Enthoven wrote that the infrastructure work took longer than expected, but the cultural shift happened faster. That inversion is notable because most adoption failures go the other direction - the tech is ready but nobody uses it.
Their explanation: every employee came in with a pre-assigned task. They weren't asked to "explore AI" or "find ways to use this." They were handed a specific problem and walked through shipping a fix. That removes the biggest barrier to adoption, which isn't fear of technology - it's not knowing where to start.
They haven't fully settled on whether to standardize on Codex or Claude Code, and they're still working on getting the agents to follow their design system consistently.
The operational detail here is in the preparation: pre-scoped tasks, real repository permissions, one-line deploys, and training that covers the full workflow rather than just how to write a prompt.