Agata retweeted
Stop reviewing agent PRs without verification! If there isn't a screenshot or screen recording, it's not worth your time.
10x more important than optimizing your CI checks.
verification is the biggest bottleneck for coding agents right now according to @addyosmani.
we @warpdotdev have put a lot of thought into building a good computer-use tool for Warp Factories. every agents gets a computer to navigate the apps they built, take screenshots and recordings, and post them back for human review.
3-4mins of activity get edited down to 15 clips stitched together with cursor clicks, mouse animations, and key presses.
we've been able to dramatically cut our time to merge and unlocked faster iteration + review cycles, especially for projects that are hard to mock out locally!
下記、似たようなアイデア考えていました。エージェンティックなe2eは追求したいところ
3. 継続的なstaging検証
- エージェントの群れがstaging環境でユーザー行動をシミュレートする
- これは、決定論的なE2Eテストでは捉えられない異常をフラグするための非決定論的なチェックを実行できる場所です
With AI writing code faster than ever, I think the end state of CICD should look more like:
1. fast PR checks on every commit
- Everything runs in <5 minutes
- Review bots go here
- Linting, typechecking, and unit tests (😱)
2. Merge queue
- All the expensive checks go here
- Integration/E2E tests are more important than ever: we need to run more of them, not less!
- If a batch fails, agent opens a PR to fix it
- Continuously merges into staging (then prod)
3. Continuous staging validation
- Agent swarm simulates user behaviour on staging environment
- This is where you can run non-deterministic checks to flag anomalies that can't be captured by a deterministic E2E test
- Catches config/stateful regressions (migration locking up DB)
コストに見合う成果はもちろん必要だけど、開発プロセスの自動化自体は自然な流れだと思う。
そもそも1年以上AIコーディングをやっていると、
依頼
↓
確認
↓
修正
↓
再確認
↓
実際に操作して確認
↓
リリース
というループ自体を自動化することも、エンジニアリングの仕事のひとつだと思えてくる。
うっかりすると「実装して」「レビューして」「修正して」「ブラウザで動作確認して」と毎回AIに同じような指示を出しているだけになりがちなので、さすがにこの往復はそろそろ自動化したいところ😇
最近、人間がコードを見ずにAIが実装・検証・自動マージまで進める世界を考えていて、GitHubのPRにちょっと限界を感じ始めている。
Diffが中心のUIだし、IssueとPRに情報が分かれているし、Conversationもノイズが多い。何が変わって、どこまで検証されているのか、直感的にわかりにくい。
要求から実装、検証、マージまでを「Change」みたいな単位でまとめて、変更内容や検証結果がぱっと理解できるUIがあるとよさそう。
PR自体は裏側の仕組みとして残してもいいけど、人間が見るUIはそろそろ別のものになってもいい気がする。
Agata retweeted
"Software development is dead" they said.
Last month a client paid me $20,000 for an app. My cost was $200 for Claude Code.
It's a kiosk mode app for android tablets in taxis. Ads for partner companies, tipping the driver, admin, backend, the whole thing. Now live in production.
I didn't write a single line of code. Still took 1 month of constant work.
The best part? The client came to me with a vibecoded version first. It was horrendous.
He paid $20k because he needed someone who has done this before, professional design, real assurance, and 12 months of support.
You can't vibe code that.
Without AI it would have taken 3 months and cost him 2-3x more.
Developers using AI will always win. Stop listening to people who say software is dead for views.
今後はよりPRの検証に使用するトークンがより大きくなる気がする。
検証はいままでは人間の目によるコードレビュー+単体テスト、実際の動作検証など保証していた部分。
人間がコードを読まないも当たり前になっているし、人間が動作検証やらないものあたりに前になるとおもう。動作検証をAIにうまくやらせるのを頑張る感じ。
ここは薄くやることも、厚くやることもできる領域だとおもうけど、網羅的かつ探索的にエージェントが受け入れテストをおこなうと、かなりのトークンが必要になるし、エージェント開発と同様に検証の仕組みの検証機構(Eval、Trace、コストなどなど)も一式必要になりそう。
Strongly agree.
A year ago, I was trying hard to get AI to write code and choose architectures that I considered best. Now I’m starting to wonder if my detailed instructions might actually be becoming a bottleneck.
Frontier models can already write solid code without much hand-holding. Maybe we’re just projecting our human preferences onto something that works differently.
Clean code evolved around how humans think and work. AI agents have different constraints — context windows, reasoning, tooling, etc.
Who knows? Maybe the best code and architecture for AI look nothing like what works best for humans.
Perhaps I’ve just spent the past year teaching AI my personal preferences 😂
And there’s a separate question: do humans really need to read code to verify correctness? Reading code has never guaranteed that a system actually works.
I think we need to rethink both how AI should structure code and how we verify its correctness.
We don’t know the answers yet, and that’s what makes this so interesting.
it's very interesting how many replies i've gotten about maintainability and clean code being still important. i would argue that that is no longer true and we all need to update our priors on what Good Code really means in 2026.
clean code has always come at a cost - performance. indirection and making code readable/understandable for a human is not free, it often comes at the cost of worse performance. we have historically been okay with this because of Moore's Law and the fact that most code is not on the hot path.
these principles also came from a time where humans maintained code. it made sense to optimize for humans then, but it makes less sense now when agents are already writing all the code. we will in fact probably discover and come up with new principles: principles of agent friendly architecture (i plan to write more thoughts about this soon)
















