Terminal agents got surprisingly good this year. Anthropic's Claude Code launched in February, OpenAI's Codex CLI got much better in August with gpt-5(thinking-high) and again in September with gpt-5-codex(high). We've been delegating bug fixes, UI features, backend updates, comprehensive testing, and even larger architectural changes to these agents at Emotion Machine. It works.

The shift from vibe coding to what Simon Willison calls vibe engineering means we can finally incorporate actual software engineering practices into terminal agent workflows—detailed planning specs, context from all stakeholders (not just developers), proper testing in deployment pipelines—while being more ambitious by running 10-20 agent sessions per person per day. But to make this work at scale, you need agents to pull from the same context pool where product discussions, design decisions, and technical specs already live. You need tighter integration between how teams plan and how agents execute.

That's what Overnight is. A way to fire coding agents from Linear issues and keep them isolated enough that they won't break things.

Here's how it works. You label an issue in Linear, Overnight spins up a Modal sandbox (cheap, isolated, disposable), clones your repo into a dedicated branch, and runs Codex or Claude Code with full codebase access. The repo clone is cached in a Modal volume and kept in sync with main, so you're not waiting on git operations every time. First pass is planning, the agent reads your issue, explores the code, writes an implementation plan and posts it back to Linear for review. You comment with changes or just say go, and it does a second run in a fresh sandbox to implement everything and open a PR. Modal sandboxes mean your secrets stay contained and agents can't do anything weird to your actual infrastructure.

The reason Linear fits well here is that it already has all the context engineers and designers are putting into issues anyway. Descriptions, comments, attached designs, links to docs, it's all there. And the mobile app is actually usable, which means you can kick off agent runs or review plans without being at your desk. There's no backend to maintain, Linear webhook events trigger Modal endpoints directly, which spawn the sandboxes, so you manage everything from Linear. We're planning to open-source this soon once we feel confident it won't eat anyone's codebase.