# A repo page for code that agents wrote > Source: > Published: 2026-10-07 14:11:04+00:00 Open any repo on GitHub and you see files, commits and people. That was the right picture when people wrote the code. In our own repo, 93.9% of the code was written by an agent. The person in the author column mostly typed a prompt and pressed enter. So the GitHub page is still correct, but it answers the wrong questions. It tells you who pushed. It does not tell you which agent wrote it, which model, from what prompt, what it cost, what the agent decided on the way, or what it left unfinished. Those are the questions you actually have now. We rebuilt the repository page in Origin around those questions. It looks like GitHub on purpose (same tabs, same file tree, same commit list), so there is nothing new to learn. Every view just carries the AI context on top. The bar at the top is the first thing GitHub cannot show: how the code splits between agents and people. Here it is Claude 93.9%, Cursor 3.8%, Codex 1.5%, human 0.8%. This is measured from capture of the actual sessions, not guessed from commit messages. The heatmap is the familiar contribution graph, but each square is coloured by which agent made the commits that day. You can see the day the team switched from writing code to directing agents. Below it is the agent scorecard. For each agent: commits, lines added, sessions, cost per commit, how much was later reverted, and how much was rejected in review. On this repo Claude made 3,821 commits at $5.18 each, Cursor 122 at $0.02, Codex 82 at $0.51. That is the table you want when someone asks whether the AI spend is worth it, or which agent to give the next task. The About panel on the right is not a static description. It is the repo’s continuation brief: what the last sessions worked on, what is in flight, and what is unfinished. More on that below. The file tree works like GitHub’s, but every folder shows how much of it AI wrote and what it cost, and every file shows the agent, the person who ran it, the cost and the age. You can find the parts of the codebase nobody on the team has really read, or the folder that ate the budget, in a few seconds. git blame tells you who committed a line. On an agent-written repo that is nearly useless, because it is the same person on every line. Origin’s blame groups lines by the prompt that wrote them. Each block says the agent and model (Claude, Opus 5), the person, and the prompt number. Hover a line and you get the prompt text itself, here “merge them, release the cli and deploy”, and the commit it landed in. The header says the file is 100% Claude across two models, written by two people, linked to PR #1977, and that 0 of 5 sessions behind it were reviewed. That last number is the one that matters in review: it tells you how much of this file a human actually looked at. Same commit list as GitHub, with the agent and model on every commit and the prompt right under the title. You can filter by agent, by model, or search the prompts themselves. “Find the commit where we asked it to change the sync interval” becomes a search, not an archaeology project. **Memory: what the agents decided, and what they left open This is the part that has no GitHub equivalent at all.** Every agent session starts from zero. It does not know what the last session tried, what it decided, or what it left half done. So it re-reads the code, re-makes the same decisions (sometimes the opposite ones), and re-discovers the same bugs. On a team with several agents and several people, this happens all day. Origin keeps a memory for each repo. After each session it writes down what was done, what was decided and why, and what is still open. It lives in the repo’s git notes (refs/notes/origin-memory), so it travels with the code and works offline. The next agent reads it before it starts, through the CLI or the MCP server, whichever agent it is. Claude can pick up where Codex stopped. The Memory tab shows that record to people. The decisions list is the useful bit: each one says what was chosen and why, and links to the session. When you wonder why the code does something odd, the answer is usually already there. The History tab shows how the memory changed over time, and you can read it as of any date. **Issues: the TODOs agents leave behind** Agents leave a lot of loose ends: “this still needs a check on Windows”, “not deployed yet”, “the old data is still wrong”. Usually that text sits in a chat log nobody opens again. In Origin, those open items become Issues on the repo automatically. They have a priority, they show which session raised them, and they close when a later session finishes the work. The [Verify] ones are checks an agent said a human should do. So the backlog of agent work is visible, and the next agent sees the same list through memory. **Why this matters now** GitHub was built for a world where the author of a commit wrote it, remembered why, and could be asked. None of that holds when agents write most of the code: The author column lies by omission. One person can run four agents on six models in a day. Without attribution you can’t tell which agent writes code that sticks, which one gets reverted, or what any of it cost. The reason is lost. The prompt and the decisions are the real source of a change. They live in chat logs on one laptop and disappear. Origin keeps them next to the lines they produced. Review can’t keep up. Nobody reads every line an agent writes. Knowing which files are AI-written and which sessions nobody reviewed tells you where to spend the review time you have. Agents forget. Every new session pays to rediscover the repo. A shared memory and an open-issues list means the next agent starts from where the last one stopped, not from zero. For an engineering lead, this is the repo page that answers the questions you get asked now: how much is AI, which agent, what did it cost, did anyone review it, what is still open. For the agents, it is the context they need to not repeat each other. Connect a GitHub or GitLab repo at getorigin.io and install the CLI. The repo page fills in from your existing history, and every new session adds to it.