>Code review is a fantastic mechanism for catching bugs and sharing knowledge
"Sharing knowledge" is one of the first phrases in the article, and highlighted as a key benefit of code review. But the loss to human-capital from this process is never examined in the post.
> Trivial reviews (typo fixes, small doc changes) cost 20 cents on average
They did around 25,000 of these runs (about 20% of total). So CF spent $5k in the period making language models run through PRs which were <10 lines long. I get that CF engineers are paid well, but the labour cost of having an intern/entry level engineer spend ~30-60s looking through these is likely close to $0.20, and that engineer builds some human-capital while they're at it.
the labour cost of having an intern/entry level engineer spend ~30-60s looking through these is likely close to $0.20
Did you do the math? Your estimate feels way off. First, I doubt an intern would process one PR in 30s. Maybe 2-3 minutes, to read 10 lines carefully looking for typos and indentation mistakes. We pay interns close to $100K these days (in a company like CloudFlare), so that's ~80c/minute. My estimate is therefore closer to $1.6 per PR. About 10X.
You are correct that there is a residual value with the intern, over time they would start learning (a little bit) about the code base.
This blog post is full of small inconsistencies that make it read like a low quality SEO piece.
> We also extract a shared context file (shared-mr-context.txt) from the coordinator's prompt and write it to disk. Sub-reviewers read this file instead of having the full MR context duplicated in each of their prompts. This was a deliberate decision, as duplicating even a moderately-sized MR context across seven concurrent reviewers would multiply our token costs by 7x.
No, it would not, because neither is the prompt of the subagent 100% of its token usage, nor will the "shared-mr-context.txt" which is then being read have a size of zero compared to the creation of this shared context.
> You don't need seven concurrent AI agents burning Opus-tier tokens to review a one-line typo fix in a README.
Yeah, well you wouldn't have anyways. Earlier in the post it says that Opus is "exclusively for the Review Coordinator".
Also this seems plain wrong. Input token caching has now idea whether you @include the file or copy the contents into the prompt. That is handled entirely by opencode and, all else being equal, has no bearing on the cache ability of a trace.
> Our cache hit rate sits at 85.7%, which saves us an estimated five figures compared to what we would pay at full input token pricing. This is partially thanks to the shared context file optimisation — sub-reviewers reading from a cached context file rather than each getting their own copy of the MR metadata, but also by using the exact same base prompts across all runs, across all merge requests.
we’ve been struggling with review throughput. this actually seems worthwhile to build at this point though i remain fairly skeptical of workflows that are agent-only, at a point it seems like the only practical solution.
we are finding lots of value in self review. its the “imagine you are doing a synchronous paired review with someone - anything that is difficult to explain, has a code smell, doesnt fit the architecture of the system around you, write a comment.” then at the end, agents do a good job of looping over PR comments.
the second thing would be a guided, educational code review tool -
there are a few attempts at this, but nothing that has a good enough interface to actually stick. organize hunks by semantic importance, spend some tokens exploring the surrounding systems, showing how new code, public apis and data model flow with the existing design, and allow a human to traverse larger PRs more quickly.
I do think Cloudflare probably institutes a similar manual review process as well. I have a handful of fairly vocal and supportive engineers I stay in contact with around https://plannotator.ai (there is an integrated code review surface that creates a feedback loop with your local agent).
> agents do a good job of looping over PR comments
This is the easy part. Most harnesses enable some sort of integration now, so you can actually create a smooth local experience around this as well - better code before it ships to more costly review or bloats PR threads.
> guided, educational code review tool
This is a bit tougher, and I find the main harness chat tends to work best. I learn better when I'm more engaged and aware of what I'm asking. It's easy to stick a code tour type of thing on a screen. It's hard to really nail the right attention and learning mechanism around it.
> Today, when an engineer at Cloudflare opens a merge request, it gets an initial pass from a coordinated smörgåsbord of AI agents.
I’d prefer to have that happen as some sort of pre commit hook, before a merge request is sent. The feedback loop might be a bit faster and the process might produce less noise this way.
It is easier to view code review results in a tool and not in a text during commit.
There is no universal standard which IDEs support for code review results (there is SARIF, but it is not supported that widely). A review result on a web page with comments from humans, is valuable.
My company has the AI review agents, and you can run them locally, but practically it’s easier to just open a merge request to have CI run the agents. Especially if you’re juggling a bunch of merge requests.
Like it or not, the "merge request" (eg, open a PR) is the Schelling point of relevant information. I expect that At scale here refers to size of software projects, and not only code velocity. Software projects of large enough size have CI configuration that don't typically fully-run on each dev machine.
Valid, but you lose the lived history that comes with the audit log of it being actual review back and forth and CI runs vs lost to a developers machine and only a relic in the commit log. I can see both sides, though.
Can you elaborate about the practical value of having the history of back and forth, in a PR or even in the commit log? In my 20ish years of experience, I can’t recall a single instance where I’ve solved something thanks to having this work-in-progress state persisted in the repo history.
It’s exclusively been the other way around where having a smaller number of larger squished commits (post merge) that’s made the project be more maintainable.
It's not about having it in the commit history. I've seen a few cases where the back and forth revealed that the AI reviewer was offering bad advice (and a few others where I suspect bad local AI advice is why people keep sending me the same category of mistake).
I have mixed feelings, but it boils down to how long it takes and / or cost.
Pre-commit hooks should be fast, as it's something you'd do (normally) a few dozen times a year. I don't believe sending a review job to a remote agent is fast, nor will waiting for a review to finish a commit be good for anyone.
CI on the other hand can be slower and runs async, it's fire and forget so you can switch tasks.
If noise is an issue, one possible solution is to create a merge request, have the tools review it, make the fixes, rewrite history like you did it perfectly the first time ("fix" commits are noise), then create a new one for human review.
Pre commit only happens on your machine though - you lose the ability to have a shared review surface where you can tag others on your team to specifically prompt discussion or verification on issues that touch their domain. When an agent points out a potential security issue with how my work ties into infrastructure, I want to just be able to tag our infra team and ask “hey is this something to worry about?” The agent, myself, and the other team member have now all contributed to a threaded discussion that is easily referencable in the future.
I think approaches like this don't need to run other than locally. Maybe integrated as pre-push hook. The system is nondeterministic, so it's at odds with the purpose of CI.
Human review is about learning and there's an implied social contract in that someone is giving you their time to make you better. It isn't necessarily necessary but replacing it with AI shows a fundamental misunderstanding of why it is part of the process.
I'd argue it's pretty much like monitoring, which certainly benefits from multiple people seeing the same stats and alerts. I agree it's at odds with CI/CD and should probably not block anything, like deterministic checks commonly do.
As a solo dev or rather nowadays more so only a decision maker / agent overseer, I came to enjoy letting my agents develop against a Gerrit repository / workflow. Dev agent pushes a CL, review agent picks it up (not just the diff, but the full repo), runs tests/reviews/review-subagents and concludes by posting a review as well as a vote. This goes back and forth with new patch sets / replies to the threads. Eventually the CL gets a +2 or whatever and I have the final call to manually submit it.
It is way slower compared to just pushing through development with one agent doing everything yolo against a normal repository, but it seems to me that the additional time is well spent (no, I don't have fancy graphs or similar analysis to prove this other than my gut feeling after looking at recent development results).
I built a more naive version for our team using Copilot and GitHub actions and it works quite well (wish I had metrics too). The team loves it.
The ROI here is so high that I don't mind using the strongest model available for the actual code review. I don't trust Sonnet and such. Just let Opus or GPT 5.5 do the whole thing and pay a bit more for less complexity.
I have about 15 or so subagents doing reviews from different perspectives (or providing some additional value, like finding agents.md files, doing confidence ranking, describing images attached to the PR, that get validated later on with Jira issue description).
I used it since about November, with large scale popularity in my company reaching in April - all that on a 300 premium requests (because they allowed starting subagents, and there was no limit how long a single request can last) - so it would cost something like $5000 and $8000 for April and May if it was API pricing. I had similar cost per review (about $0.90) with Opus 4.6 and help from Sonnet and Haiku for simpler tasks. It did about 4000 reviews during the last 2 months.
And starting in June, it will be dead because it will be API pricing and for $30 (or $19 since September) it will do just few reviews.
Every iteration something can be found. How many times do you iterate e.g. on performance - use optimized struct, oh, you can change the architecture etc.? At that point one can just have a while loop for the agents to make changes until no comments left.
> One of the operational headaches we didn’t predict was that large, advanced models like Claude Opus 4.7 or GPT-5.4 can sometimes spend quite a while thinking through a problem, and to our users this can make it look exactly like a hung job.
I had the same problem in my recursive agent harness. It would always come back, but it could sometimes take up to 10 minutes depending. I fixed this by adding a required "purpose" argument to every tool and call/return event. As the recursive evaluation proceeds, every single thing that happens streams incremental purpose text to the user's browser (also using the magic of JSONL for this). The incremental progress events contain the purpose and a detail section (tool arg JSON) that the user can expand/collapse.
You can do basically the same thing as cloudflare except as a skill you run in your local harness. If you're going through the motions with PRs and are familiar with actions, you can have it run in a github action instead. But this is basically just a skill. The Claude code review skill is a simple version of exactly this.