If you've been to our homepage today, a small green character may have popped up from behind a terminal to tell you, a little awkwardly, that Coderbuds works with Amp. That's Puck, Amp's assistant. This is the longer version of what he was trying to say.
#We Build Coderbuds in Amp
We use Amp every day. In the 30 days to 20 November, 103 of the 175 pull requests merged into Coderbuds came out of an Amp thread, by the Amp-Thread-ID trailer Amp puts on its commits.
So supporting Amp properly isn't a partnership line for us. It's how we find out whether Coderbuds tells the truth about the way we actually work.
#Why We're Aligned
Amp's team wrote down how they work, and one line in it could be Coderbuds' brief:
"A thread ends when the change has landed and we know it works, not when the code is written."
That's the question Coderbuds exists to answer. It isn't how many lines someone wrote, or how many pull requests a person opened. It's whether a change reached production and held up there. Amp puts the agent in every step from bug report to rollout. Coderbuds measures the result of all those steps together: lead time from the change to the deploy, deploys that failed, changes that got reverted.
We also agree about where people belong. Amp's position is that agents can now catch the bugs and little mistakes, so humans don't need to review every change. Ours is that a silent LGTM is worse than no review: spend people where a mistake costs something, and let the routine changes go. Those are two ways of saying the same thing. Mandatory review of everything stopped working once agents multiplied the changes.
#The Part Amp Does Differently
Amp ships without pull requests by default. From Amp's docs on shipping:
"By default, Amp uses trunk-based development. This means there is one shared branch and every change lands on it directly. The agent commits any uncommitted changes, rebases onto the latest base branch, runs the tests, and pushes straight to the base branch."
And Amp's own team: "Our core product team pushes to main… It's no longer necessary to open PRs and get reviews to catch bugs or little mistakes."
So any team that turns on Amp and leaves the project settings alone pushes straight to main.
Coderbuds was built on the pull request. Lead time started at merge, AI attribution keyed on a pull request number, the review policy graded pull requests. A change pushed straight to the branch had none of that. Coderbuds saw the deploy and nothing about the change inside it. Worse, a deploy with no linked pull requests could borrow the lead times of unrelated ones merged before it, so the number wasn't just incomplete, it was wrong.
That's no good for a tool that says it measures how your team ships, when the most popular way to ship with one of the best agents skips the thing it was measuring.
#What Coderbuds Does With a Direct Push Now
The unit Coderbuds measures is now the landed change: a change that reached your default branch and production, whatever route it took. A pull request is one route. Pushing straight to the branch is another.
- Every change is recorded, whatever its route. When a deploy goes out, Coderbuds walks the commits it carried on the branch's first-parent line. A merge or squash commit is its pull request's change. A commit that a merged pull request contains is a rebase-merged one. Anything else was pushed straight to the branch. Direct commits from one Amp thread count as one change, and link back to the thread. If GitHub can't tell us whether a commit came through a pull request, we record nothing for that deploy rather than guess.
- Lead time counts direct changes. A direct change is timed from when it landed to the deploy that carried it, next to merged pull requests. For a direct change, the push is the merge. A deploy that carried only direct changes no longer borrows lead times from pull requests it didn't contain.
- The Review tile shows the direct route. Beside the pull requests, it shows how many changes were pushed straight to the branch, how many of those needed no review under your policy, how many went round it, and how many held, meaning not reverted and their deploy not marked a production failure. The weekly review-policy post in Slack carries the same line. You can open the list of direct changes to see what each one was, where it landed, whether it held, and the Amp thread it came from.
- People are never named. The direct route is counted for the team. Nobody gets a "shipped without review" leaderboard.
Coderbuds takes no position on pull requests versus pushing to main. It shows whether each route holds up for your team, and lets the evidence argue.
#How the Agent Decides: Ship or Pull Request
The review policy is what makes this safe rather than just visible. Before an agent pushes, it calls Coderbuds' assess-change-fit tool with the diff, and gets back the review tier your policy gives that change.
- If the tier is no review needed, say a docs fix or a test, shipping straight to main is fine. Press Ship.
- If the tier asks for any review, an AI review included, the agent pushes a branch and opens a pull request instead, where that review happens and is recorded.
After a direct push, the agent reports it with report-ai-usage, giving the commit SHAs it landed, so the change is attributed to Amp rather than guessed from the shape of the diff.
That makes your team's own policy the agent's answer to "Ship or PR?". Coderbuds can't stop a push, and doesn't try. The instruction is guidance, and the count of direct changes that went round the policy is how you find out it was ignored.
#Two Honest Caveats
You need deploy tracking. Coderbuds finds direct changes by walking what each deploy carried, so a team that doesn't send us deploys won't see them yet. Deployment tracking works from GitHub Actions or Bitbucket Pipelines.
Our own repository is a weak test of the numbers. We still mostly use pull requests. In the 30 days to 15 November, 111 of the 112 changes on our main branch came through one. That proves the plumbing works. Comparing the routes properly needs a team that pushes to main as a habit, which is exactly the team on Amp's default Ship.
#Connect Coderbuds to Amp
Add Coderbuds as a remote MCP server, so ampcode.com holds the sign-in and every thread can use it, including threads in orbs:
1amp mcp remote --personal add coderbuds https://coderbuds.com/mcp/insights --auth oauth
2amp mcp remote --personal login coderbuds
A workspace admin can add it once for every member instead; the Amp section of our MCP docs has both. Then ask the agent to set the repository up for Coderbuds. That adds two rules to your AGENTS.md, so the agent checks before it ships and reports after.
Then press Ship. Puck would want you to.