Zag推出Apple生態AI審核Agent，解決程式碼代理氾濫審核難題
Zag推出Apple生態AI審核Agent，解決程式碼代理氾濫審核難題。
Zag是YC F25新創，為iOS、iPadOS與macOS開發者打造AI審核Agent，在真實Apple開發環境中執行，包括Apple Silicon、Xcode、模擬器與完整工具鏈。開發者以TypeScript或Swift描述Agent，處理程式碼審核、QA、安全性或App Store合規檢查，並在每個PR上自動運行。
審核必要性
隨著程式碼Agent產生人類無法審核的變更量，每個團隊將需自建客製審核Agent，符合自家程式庫、標準與行為規範。唯一強制工程文化的途徑，是在主分支每個變更提案上部署審核Agent。靜態AI程式碼審核僅提供內嵌註解，需過多人類介入，團隊難以持續；Zag Agent實際建置、測試與運行App，透過點擊流程或macOS套件的紅隊攻擊產生信心。
Agent設計與執行
Zag審核Agent嵌入程式庫，由團隊維護，以TypeScript或Swift描述。主要工作單位為「Task」，定義環境與內部邏輯，可依賴其他Task。
- Zag SDK簡化在Task中運行Claude Code與Codex，存取模擬器、Xcode與macOS資源。
- 內建工具快速啟用QA、安全、App Store合規等常見檢查，在Zag預熱運算上數秒完成。
每個Task在「session」中運行，即保留Apple Silicon的macOS環境，數秒啟動，提供耐久、可重播串流；Claude Code或使用者可訂閱觀察Agent運作。
定價與整合
Zag認為價值在Agent層而非運算利潤，故macOS執行器按AWS成本計價，無排隊的專屬Apple Silicon僅$0.015/分鐘；擴展至自有裸機叢集時，節省全數轉嫁。若已有GitHub Actions或Cirrus工作流程，一行變更即可指向Zag運算，即便不使用審核Agent。
團隊背景與啟用
Zag由Google與Tesla團隊打造，靈感來自Google驗證巨型內部monorepo的系統。目前開放iOS、iPadOS、macOS團隊早期存取，預約demo（zag.dev/book）即有專屬工程師嵌入團隊，在每個PR上設定可信審核Agent，加速程式碼交付。
Introducing Zag
— Zag (YC F25) (@zagdotdev) April 17, 2026
AI review agents for iOS, iPadOS, and macOS developers.
Zag runs agents in real Apple developer environments with Apple Silicon, Xcode, simulators, and your complete toolchain.
Describe agents in TypeScript or Swift that do code review, QA, security, or App… pic.twitter.com/yeTvM0FJnL
As coding agents produce more changes than humans can review, every team will build and maintain review agents tailored to their codebase, their standards, and the behaviors they care about.
— Zag (YC F25) (@zagdotdev) April 17, 2026
The only way to enforce engineering culture at the scale of coding agents is to deploy…
Static AI code review with in-line comments requires more human-in-loop than teams can sustain.
— Zag (YC F25) (@zagdotdev) April 17, 2026
Zag agents actually build, test, and run your app to produce confidence, whether that's clicking through your app or red-teaming your macOS package.
Zag review agents live inside your codebase, are maintained by your team, and can be described in TypeScript or Swift.
— Zag (YC F25) (@zagdotdev) April 17, 2026
The primary unit of work in Zag is a Task.
— Zag (YC F25) (@zagdotdev) April 17, 2026
A Task describes an environment and the logic that runs inside it. Tasks can depend on other Tasks.
Zag SDK makes it easy to run Claude Code and Codex inside a Task, with access to simulators, Xcode, and anything else on macOS.
Use… pic.twitter.com/QKk8oHHdif
Every Task runs in a session. A session is a macOS environment on reserved Apple Silicon that boots in seconds.
— Zag (YC F25) (@zagdotdev) April 17, 2026
Each session exposes a durable, replayable stream. Claude Code can subscribe to a session and watch the agent work. So can you. pic.twitter.com/SLcrnCIC7T
We believe the value in this category is captured at the agent layer, not in compute margins.
— Zag (YC F25) (@zagdotdev) April 17, 2026
So our macOS runners are priced at AWS cost, and we provision dedicated Apple Silicon with no queue at $0.015/minute. When we scale to our own bare-metal fleet, the savings pass…
Zag is built by a team from @Google and @Tesla, inspired by the system Google uses to validate code in its massive internal monorepo.
— Zag (YC F25) (@zagdotdev) April 17, 2026
Book a demo this week and a dedicated engineer will embed with your team to set up review agents you can trust on every PR.
Today, we're opening early access to teams building for iOS, iPadOS, and macOS.
— Zag (YC F25) (@zagdotdev) April 17, 2026
If you'd like to ship code faster with custom review agents, book a demo at https://t.co/Azkq6js8fC
— Zag (YC F25) (@zagdotdev) April 17, 2026