In January, I pasted a blog post into a coding agent. Nine months later, OpenInspect is running in production at companies from 5 to 500+ engineers.
It's time to hand it off. https://x.com/i/article/2116233023104704512
I'm handing off OpenInspect. Here's what I learned
Why I'm stepping away
I'm stepping away from OpenInspect on November 1st, nine months after the first commit. I'm starting a new full-time role (stay tuned), and it would be otherwise difficult to do both.
The repo is moving to the open-inspect GitHub organization, where a handful of maintainers will take it from here. The license stays MIT and no changes for anyone already running it.
Why OpenInspect started
OpenInspect came about to be as an implementation of Ramp's published Why We Built Our Own Background Agent. The post came just after the Christmas mania, where everyone discovered Opus 4.5 and GPT-5.2 could handle longer-horizon tasks. As a result, everyone's machines were folding under parallel agents and worktrees and the term "software factory" was nowhere to be found.
Claude web and codex had early cloud agents, but both solutions weren't designed for a company environment.
The post highlighted an internal company agent infrastructure and its benefits, matching what I was seeing and building within my clients organization.
The post ends with, "We think anyone should be able to build this... Paste the link to this post into a coding agent and let it begin building."
So I did and got to building...
45k lines of code later, the first commit landed and the project was open-sourced. Slack app, sandboxes, terraform and all in nearly one-shot.
Having seen it first hand and now with confirmation from the post, I felt these background agent systems would become critical infrastructure within every company. Surveying the market, there wasn't anything of quality that was open-source and I felt every company would need to customize it into their environment.
The hypothesis was straight forward: Give teams a high qualty background agent system that they could deploy into their own infrastructure and customize on top of. Rather than try to productize it and max-extract per seat pricing, sell FDE services to help the orgs scale it.
Impact
OpenInspect changed the way we worked
In nine months, OpenInspect went from a claude one-shot to being in production at companies from five to 500+ engineers. The repo has 3,400 stars, 500 forks, and 1900 merged PRs from 50+ contributors. By August, more than 100,000 sandboxes had launched across deployments I worked on, and likely many more across others I didn't. Many teams were running tens of thousands of dollars per month through the system.
The community jumped in as well. Several of the big integrations came from other people: the Linear integration, Gitlab support, three of the seven sandbox backends.
I talked about it, a lot! On Latent Space, at the Background Agents Summit, on Modal's livestream, and several blog posts with 100k+ views on X. Attention came in bursts and mostly lined up with content rather than releases.
Things that were correct
It became critical infra
Teams that deployed it came to love it and it quickly spread throughout the organization. Once permeated, it dominated anywhere from 30-70% of a companies pull requests. Bringing the agent to wherever the work was happening and adapt the system into the org played a huge part in adoption
The control plane is the product
The harness wars are mostly over and more of a vim vs emacs debate at this point. Consensus has mostly landed that they are a small piece in a much larger system and interchangable. Once I added an interface for it, Claude Agent SDK landed the same day, with Codex a few weeks behind it. The harness adapters make up roughly 3% of the production code, the control plane is nearly half.
The compute infra was swappable too. With my own deployment, I switched providers multiple times.
Non-technicals are building
When non-technical teams no longer need to worry about Docker and other developer goo, the barrier to contribution drops. Across every OpenInspect deployment, I saw non-technicals contributing code. Everyone from the CEO down to the marketing team. Simplified tooling allows anyone to be a builder.
A factory becomes self-improving
OpenInspect is built with OpenInspect. More than 40% of all commits came from its own sessions. 70% in October.
Things that were challenging
Harness in the box vs out of the box
The agent can run inside the sandbox, or outside it with tool calls executing within it. Running multiple different harness makes the agent-in-the-box discussion much more difficult. All of the major lab harnesses are designed to run in the box and running them outside require hacks to support and you lose functionality.
Sandbox Providers Vary Widely
Supporting seven different sandbox providers gave insight into how differently providers implement the same features. Some snapshots take forever. Some shutdown the sandbox when you take a snapshot (lol). Some had secrets egress proxies, others didn't.
Towards the end, and with some influence :D, the major players have mostly converged on their offerings and functionality.
Org Challenges
Once the system was deployed, it often came with organization challenges. Everything from company integration, spend management, adoption, quality and more. The system is a small piece in a much larger "AI-native" organization problem. Working across multiple orgs, these were some of the more fun and challenging problems to tackle and led to some functionality like Managed Skills.
Future
Outside of the project, I think everyone will end up running an agent control plane and the larger company agent infrastructure. Whether you build or buy, agents writing the majority of our software and workflows is here to stay.
Invest the time and resources to get ahead of it.
Thank you
To the 50+ contributors, everyone who filed an issue or helped answer someone elses, every team running it in prod and giving feedback
And to the new maintainers carrying the torch on. It's yours now!