Key Takeaways

  • Quinn Slack argues that AI agents operating inside reproducible cloud sandboxes make traditional CI/CD pipelines and GitHub pull requests obsolete.
  • Ephemeral sandbox platforms like E2B and Daytona let developers spin up over 100 environments a day, allowing agents to execute and fix test suites locally before pushing.
  • Swyx defends deterministic CI, arguing that AI agents cannot be trusted to run exhaustive tests across massive codebases due to context window limits.
  • Slack notes that legacy developer tools fade quietly rather than abruptly: teams stop opening GitHub issues and PRs, using GitHub only when GitHub Actions breaks.

The Disagreement

Software engineering workflows have centered on GitHub pull requests and centralized CI/CD test runners for over a decade. Quinn Slack, co-founder of AMP and former CEO of Sourcegraph, thinks that era is over. He believes AI coding agents running in isolated cloud sandboxes eliminate the need for traditional staging and verification pipelines.

“These things don't die with a bang. They die with a whimper,” Slack explains. “We just don't even think of GitHub anymore. We're not using issues. We're not using pull requests. We're barely on GitHub Actions and we're looking to get off of that.”

Slack points out that when an agent writes code inside an identical team sandbox, it runs the entire test suite to verify its own work. If a test fails or flakes, the agent fixes it on the spot before saving. Slack asks a simple question: “If my agent has run the tests for its own verification, why do I need CI?”

Swyx pushes back on removing the safety rail. His counterargument rests on the limits of agent attention: “You don't trust the agent to be exhaustive. It'll run what's in its context, but if you have a large project, you don't know that it's run everything that it should run. And so CI is just like a deterministic stage of like, hey, just run it on all these tests.”

Who's Right (and When They're Wrong)

Slack is right for early-stage teams, greenfield products, and modular microservices. When your test suite finishes in under two minutes, running full verification inside an agent sandbox (like E2B or Daytona) saves massive cycle time. Waiting twenty minutes for a centralized GitHub Actions queue to tell you about a syntax error is dead weight.

Swyx is right for enterprise monoliths, strict compliance regimes, and multi-service dependencies. AI agents optimize for the files in their active context. They will frequently skip integration tests in distant directories to save tokens or reduce latency. Until agents possess infinite context and deterministic guarantees, you still need an unskippable central gate that runs every end-to-end test before production deployments.

Centralized CI will not disappear overnight, but its role has changed. Instead of serving as the primary feedback loop for engineers, CI is shrinking into a final sanity check while ephemeral sandboxes take over actual development.

What to Do With This

Audit your team's developer loop this week. Count how many minutes your engineers spend waiting on GitHub Actions builds versus running tests directly inside local or cloud sandboxes. If your build queue takes longer than five minutes for routine pull requests, shift test execution into a pre-push agent sandbox script so failures are resolved before code ever leaves the developer's branch.