Race coding agents on tracker issues and merge the best pull request

Serhii Troian describes how the team behind Origin, an issue tracker for agent work, now runs its own tickets. Issues from GitHub, GitLab, Jira, Dependabot alerts and TODOs that agents left in the repo's memory notes land in one list. A button hands an issue to a coding agent CLI on the developer's own machine: the prompt is built from the issue text plus the repo's memory notes, and the agent runs headless in a fresh git worktree on its own branch. Before starting, the agent reads the repo in read-only mode and either writes a short brief or posts up to three clarifying questions on the issue. When the run exits cleanly the tool pushes the branch and opens a PR, but never merges or force-pushes. For uncertain tasks the same issue is given to two to four agents at once, each on its own branch, and the team merges the best PR. Runs stop after 120 minutes.

On a Dependabot alert for source-map-js, Claude Code and Codex both opened PRs; Claude's was merged at a cost of $0.50, while Codex's was closed after costing $0.96 and using about ten times the tokens. On a TODO that Claude Code had left in the repo memory, Antigravity failed its first attempt and on the second worked for 30 minutes and opened a PR of +846 and -70 lines across 9 files. Early on, a signed-out Claude Code exited in a second and the run still looked completed, which is why sign-in is now checked before any run starts.

This is a vendor post about the author's own product, and the numbers come from their own dogfooding on a few issues, not a broader comparison. Each agent runs in its narrowest unattended mode (for example Claude's auto permission mode or Codex's workspace-write sandbox) rather than a blanket permission bypass. Antigravity is not in our agent list, so only Claude Code and Codex are tagged.