Brian Douglas

Bdougie (Brian Douglas) is the Oakland-based co-founder and CEO of Paper Compute. He builds tools that preserve the exchanges between engineers and their coding agents, including the moment a teammate explains why a proposed change duplicates an existing service. His earlier company, OpenSauced, helped developers find people and projects through the code they contributed.

How a pull-request tracker became OpenSauced

Bdougie graduated with a finance degree in 2008 and went into sales. He learned Ruby on Rails to build Chuych, an app for finding churches. At a Florida Ruby conference, he showed it to two developers he knew from a local user group. They encouraged him to apply for a job, and by 2014 he was working at IZEA. When he later needed a Node.js bot to invite people to a Slack community for people of color in tech, he knew enough Ruby to build an app but too little JavaScript to get the bot running. Its author, Chad Scira, helped him work it out over email. Bdougie remembered the welcome Chad gave him. One way he extended that welcome to others was to introduce readers to emerging engineers through his Awesome Black Developers list.

As a front-end engineer and Netlify’s third employee, Bdougie attached deploy previews to his open-source pull requests so maintainers could try his changes. Keeping up with the replies was harder. He built OpenSauced for himself, using GitHub Issues to track repositories and conversations that needed his attention. After he showed it at GitHub Universe in 2017, GitHub recruited him. He joined in 2018 and went on to lead developer advocacy. In 2020, viewers watching him build OpenSauced on a livestream asked for logins so they could use it on their own projects. He added accounts and later made OpenSauced a company.

OpenSauced helped answer a question that lifetime commit counts handle poorly. Who works on this project now? The original author might still top the list years after leaving. An Insight Page showed the previous 30 days of contributions across related repositories, while a contributor could highlight a particular pull request and explain the problem it solved. Bdougie also encouraged newcomers to look for growing projects that could use a second or third contributor. Asking a maintainer how to get started and writing down the answer in a missing CONTRIBUTING.md could itself be a first contribution. He called this open-source hospitality.

The Linux Foundation acquired OpenSauced in 2024. On Open Source Ready, the podcast he co-hosts with John McBride, Bdougie said he would manage developer experience for the LFX platform and John would lead AI and infrastructure.

What survives an agent session

After his work at the Cloud Native Computing Foundation, Bdougie joined Continue and helped build its cloud coding agent. A run could last hours. If it failed overnight, the engineer investigating it needed a way to reconstruct what the agent had done. Bdougie left Continue and co-founded Paper Compute with John. Their team built stereOS, a Linux system that runs agents in isolated environments, and Tapes, an open-source proxy that records requests and responses between an agent and its model. Bdougie built Sweeper to put both to work on code maintenance. It groups lint errors by file, assigns the groups to agents working in parallel, and keeps a record of each agent’s session.

Those records changed how Bdougie handled a recurring Kafka setup problem. His Pokémon agent and Sweeper both needed to publish events to Confluent Cloud. Twice in one week he had to diagnose why events weren’t arriving. The sessions preserved details that were easy to forget, including the need to create a topic before publishing to it and a configuration field that expected the name of an environment variable where he had supplied the credential itself. He used the recordings to generate a Confluent Cloud setup skill that paired each symptom with its cause and fix.

For code review, the part of a session Bdougie wanted to preserve was a colleague’s objection. He had used Rust for side projects and found himself accepting generated code that compiled even when it departed from the team’s API conventions. His teammate Matt often caught those choices while working with an agent in the terminal. Bdougie labeled the exchanges where Matt corrected Go or Rust code and used them to train a model for Twin Review, a prototype that could comment on Bdougie’s pull requests using examples of Matt’s feedback. He kept the prototype private because its training data was sensitive. For teams with a recurring correction, Bdougie recommends starting with a written skill and setting aside other sessions to test whether the agent applies it. That gives the team a way to judge what the instructions accomplish before spending compute on model training.