Reason

Reason is an autonomous AI software engineer that connects to GitHub, takes assigned issues, edits code, writes tests and returns pull requests with recordings and passing checks attached. It also answers engineering requests in team chat and runs on Linux and Mac desktop environments.

Reason positions itself as a coding agent first and everything else second. The product is an autonomous software engineer that picks up a task, works through it in a repo, and hands back a pull request with proof attached. Its homepage demonstrates this across two scenarios: fixing a broken double boost in a small space game and building an AI spend dashboard from a billing export. In the bug case the agent traces the failure to a thrust check that resets the boost charge every frame, edits the relevant files, adds a regression test, then runs the game route to confirm all eight checks pass before merging. In practice the workflow is built around GitHub. An issue is assigned, the agent reads the codebase, makes the edits, writes tests, and opens a pull request. The difference from a generic assistant is the emphasis on evidence: the homepage shows the agent running the app on its own computer, recording a playtest, and attaching the recording to the pull request, along with a summary of what changed and why. Reviewers see a diff across source files, a changelog entry and passing checks rather than a block of suggested code pasted into a chat window. Reason also lives inside chat. In the shared channel shown on the site, a teammate asks for a quarterly reliability review and the agent pulls incident and uptime data, produces a twelve slide deck and two CSV exports, and replies in thread. That makes it useful beyond pure engineering work, but the company is explicit that coding is the focus and other abilities are a byproduct. The site compares its performance to Codex, claiming it outperforms that agent while being 22 percent more cost efficient. It suits engineering teams that already run issue driven workflows on GitHub and want routine fixes, tests and small internal tools handled without a person driving each step. Expect a review-heavy process: the agent proposes, a human merges. Downloads are offered for Linux and Mac, with desktop, terminal and review surfaces shown in the product screenshots.