Most of our tickets end the same way now. Someone copies the issue text into Claude Code or Codex, adds “look at the memory notes first”, waits, pushes a branch, opens a PR and pastes the link back into the ticket. The work is done by the agent. The glue around it is done by hand, and nothing records which run produced which PR, what it cost, or why the first try failed.
So we moved the glue into Origin. An issue now has a Start with button. It hands the issue to an agent that runs on your own machine, in its own worktree, and the run is recorded on the issue: who ran it, which agent and model, how long, what it cost, how many lines, and whether its PR was merged. This post shows how it looks and what it does. A separate write-up covers how we built it.
One list, every source
Issues come from five places and land in one list: GitHub and GitLab issues (two-way sync), Jira, issues created in Origin, security alerts from Dependabot and code scanning, and TODOs that agents left in the repo’s memory notes. The list sorts by priority and filters by repo, source and assignee. Each row shows the source, labels, priority and, once an agent has worked on it, the number of sessions and the AI cost.
The host stays the owner of what it owns. Title, body, labels, assignees and open/closed come from GitHub or GitLab, and Origin writes assignees back. Origin owns the states the host does not have: in progress, in review and blocked.
Start with an agent
The menu lists the agents Origin can launch headless: Claude Code, Codex, Gemini CLI, Cursor, Copilot CLI and Antigravity. Pick one and you get the Start dialog.
The prompt is built from the issue and the repo’s memory (here 4,418 characters, 10 lines of it memory), and you can edit it before you start. Below it is the exact command the button runs, so you can run the same thing from a terminal:
origin issue start OR-44 --agent claude --background
Without Run in the background, Origin opens a terminal in your checkout with the agent and the prompt in it, and you work as usual. With it, there is no terminal. The agent runs headless in a new worktree on branch issue/OR-44-…, so your own checkout can stay dirty. When the agent exits cleanly, Origin pushes the branch and opens a PR. It never merges and never force-pushes.
The run happens on your machine, with your agent CLI and your agent’s account. Origin’s server never runs code. The button opens a getorigin:// link that carries only the repo and issue ids; the Origin CLI on your machine reads the issue with your own key and starts the agent. The full log stays in ~/.origin/agent-runs/. The server gets the status, the exit code and the last few log lines.
Before it starts: is the agent signed in, and is the task clear?
A signed-out Claude Code exits in a second with “Not logged in”, and for a while that run still looked completed. Now a background run first checks that the agent is signed in. If it is not, nothing starts, no worktree is made, and the issue shows a failed run with the reason. The Agents page shows the same check per machine; above, the last background Claude run on this MacBook could not sign in.
The second check is about the task. A one-line issue leaves the agent guessing, and a background agent has nobody to ask. So before a background run, Claude or Codex reads the repo in a read-only mode, without hooks and without saving a session, and answers one question: is this clear enough to do? If yes, it writes a short brief (files and approach) into the prompt. If not, it asks up to three questions. They appear on the issue, nothing starts, and you answer on the page or with:
origin issue answer OR-44 "Keep v1 readable; consumers must handle both"
A run can also stop and ask halfway. When you answer, the next start resumes the same agent session (claude --resume, codex exec resume) if that session is on this computer, so the agent keeps what it already read. Otherwise it starts fresh with every answer in the prompt.
Every attempt stays on the issue
This issue was not written by a person. Claude Code left it as a TODO in the repo’s memory during an earlier session (“the header’s tokens, cost and tools chips still come from the old session totals”), and Origin turned it into an issue. Antigravity took it twice. The first run failed and the table says so; before we recorded runs, a run that died before the agent’s first hook fired left no trace at all. The second run worked for 30 minutes and opened PR #2487 with +846 −70 lines in 9 files. The issue moved to In review on its own.
Each row links to the captured session, so you can read every prompt, tool call and diff of that run. Status follows the work: start or claim → in progress, PR open → in review, PR merged → closed, PR closed unmerged → back to in progress. An agent can also end with a line like [Origin: Outcome] blocked — which schema version?, and the issue goes to blocked with that question on it.
Race two agents on the same issue
If you are not sure which agent is best for a task, give it to two to four of them. Each runs in the background on its own branch from the same base, and each one that finishes cleanly opens its own PR. Merge the best one, and the issue closes.
This one was a Dependabot alert for source-map-js. Origin turned it into an issue with acceptance criteria (the lockfile no longer resolves a vulnerable version, tests pass). Claude’s PR was merged at $0.50. Codex’s PR was closed; it cost $0.96 and used about ten times the tokens. Over a few dozen issues, that table tells you which agent to give which kind of work. For closed issues the page also gives a origin benchmark replay-tasks command, so you can replay the issue on other agents later and grade them against the tests the merged PR added.
Agents that pick up work themselves
The Start button is one way in. The other is the queue. Origin’s MCP server gives agents list_tasks, take_task, finish_task and release_task. An agent in a session can take the next ready issue, and two agents can never take the same one: a claim is a single conditional database write, so exactly one of them wins and the others get “claimed”. A claim is a lease (30 minutes by default) renewed by the session’s heartbeat. If the agent dies, the lease runs out and the task is free again, without a cleanup job. Issues with open dependencies are not served.
What it will not do
Run anything on Origin’s servers. The agent runs on your machine, with your installed CLI and your account.
Use a blanket permission bypass. Each agent gets the narrowest unattended mode it has: Claude’s --permission-mode auto, Codex’s workspace-write sandbox, a per-run Antigravity project allowed only git and read-only commands with no network, a scoped .cursor/cli.json.
Merge, or force-push. It opens a PR and stops.
Run forever. Background runs stop after 120 minutes by default, together with every process the agent started.
Send the whole log to the server. Only the status and the last lines.
Try it
You need the Origin CLI and at least one agent CLI signed in. Run origin url-handler install once so the Start button can reach your machine, then open Issues and press Start with. From a terminal it is origin issue start --agent codex --background, origin issue race --agents claude,codex and origin issue runs to see what is running.
Top comments (0)