Quinn Slack AMP Killed Mandatory Code Review

Latent Space caught Quinn Slack AMP at the moment the team killed mandatory review before merging to main.

StartupHub data

Companies working on this

Profiles of the companies named in this story, with funding and a one-liner from our database.

Hugging Face

$4.5B

Hugging Face is the leading AI community and platform for machine learning collaboration, enabling developers to build, share, and deploy models, datasets, and applications.

What was demoed was Orbis, AMP running agents remotely in E2B sandboxes with 100 parallel tasks, live dev servers and desktop access in the portal. The affected system is AMP itself, still pushed to GitHub today, increasingly to private Pier forges tomorrow. No remote exploit is needed. Any teammate with push access can ship direct to main.

Slack is Sourcegraph's former CEO who spun AMP out from Sourcegraph with 20 people, every investor and employee holding both companies. That history matters because the model assumes a high-trust, owner operated team, not a 200-person queue.

The industry context is crowded. Devon proved early that remote agents could work, and now Conductor, Opencode, Claude and Cursor are all pushing cloud execution. Even media tooling like Overclip backed by a16z came up as Slack described how agents edit everything else.

How the Quinn Slack AMP no-review workflow actually works

Think highway without toll booths but with fast tow trucks. The team optimizes for mean time to recovery over mean time to failure, aiming to ship a fix 15 minutes after a log or customer report instead of waiting days.

Scope replaces review. Instead of a local CLI agent inheriting every credential on a developer laptop, which Slack noted has escaped sandboxes and wandered into prod consoles at customers including a Hugging Face style breakout, AMP issues OIDC tokens for 30 minutes of read-only logs or database access inside a repeatable isolated environment.

Why this matters and what is not fixed

Agents now run tests in the same environment for everyone, and Slack argued that beats flaky CI gates if you fix flakes and let the agent rerun what it touched. The gap is that this only holds with senior owners and deterministic tests.

For most builders review is also the sharing and security control. Copy the removal without the compensations and you trade queue time for audit blindness. AMP itself has not fully removed CI, and still uses GitHub Actions as the catch-all for what the agent skipped in context.

The tell was a 12-hour flight to Munich. Slack let an agent ship a multi-phase migration from single to multi-workspace, watch logs and metrics after each stage, and auto-rollback if errors spiked.

When you realize you have not run localhost in two weeks, local dev and mandatory review do not end with an announcement. They just fade.