Zag

About Zag

Coding agents are changing how software gets written. Tools like Claude Code and Codex can author changes faster than any team can review them, and the gap is only getting wider. The bottleneck in software development is shifting from writing code to validating it.

We think every serious engineering team will need review agents: custom agents, tailored to their codebase, their standards, and the specific behaviors they care about, running on every proposed change to their main branch. Static AI code review with inline comments won't scale. The kind of confidence that lets teams merge as fast as they author comes from agents that actually build, test, and run the code.

As human engineering work moves up a layer of abstraction, we believe a tech lead's primary job will be tuning better review agents. Not reading diffs line by line, but shaping the automated judgment that enforces their team's engineering culture at the scale of AI-authored code.

Zag is the platform for building and running those review agents. You describe agents in TypeScript or Swift as composable Tasks. Zag runs them in real sandboxed environments with real build tools, real test suites, and real running applications on every PR.

Today, we've opened early access to teams building for Apple platforms. Zag provisions dedicated Apple Silicon with Xcode, simulators, and the full native toolchain, so review agents can do what humans do: build the project, boot the app, and verify that it works.

How We Got Here

Zag is inspired by the system Google uses to validate code across its internal monorepo, where every proposed change is reviewed by both humans and automated agents before it can land. We believe that model, built for a world where machines participate in the review loop as first-class citizens, is the right shape for what's coming.

The Team

Zag is built by a small team from Google and Tesla based in San Francisco. We're backed by Y Combinator and building in the open at zag.dev.

Working at Zag

We're early. The product is live, the infrastructure works, and the first teams are onboarding. What we're building now will define the platform for a long time. You'll work directly with the founders, ship to production daily, and have real ownership over the systems you build.

We care about craft, clear thinking, and building something that engineers actually trust. If the idea of shaping how AI-authored code gets validated, and doing it by building beautiful, reliable infrastructure, sounds like your kind of problem, we'd love to talk.

Zag is a platform for building and running AI review agents that validate code changes in real sandboxed environments. The core technical problems we're solving:

Agent orchestration. Zag's primary unit of work is a Task: a composable, dependency-aware unit that describes an environment and the logic that runs inside it. Tasks can spin up Claude Code, Codex, or custom agents with access to real build tools, simulators, and full operating systems. The SDK is TypeScript-native, with Swift support for Apple platform work.

Sandboxed compute. Every Task runs in a session: an ephemeral macOS environment on dedicated Apple Silicon that boots in seconds from pre-warmed snapshots. Sessions expose durable, replayable streams so agents (and humans) can watch work happen in real time. We manage the full lifecycle of these environments, from provisioning to teardown.

Code review as infrastructure. Zag's review model descends from Google's internal code review system, where every proposed change is validated by both humans and automated agents before it lands. We're rebuilding that model for a world where coding agents produce more changes than humans can read, and review agents need to actually build, test, and run the code to produce useful signal.

Stack. TypeScript, Swift, Go. Gerrit (forked). Tart for macOS virtualization. Apple Silicon bare metal. GitHub and Git integrations.