7.5 issues per PR.
4 Sev-1 incidents.
The slopocalypse is here … and code review is far from a solved problem.
At enterprise scale, senior engineers are already drowning in an avalanche of PRs, driven by four factors:
Top-down organizational pressure for everyone to use more AI and ship faster. Like every other tool, AI is a multiplier of operator skill, and simply “using” AI doesn’t lead to better results.
Formerly-non-engineers making changes without implicit or explicit guardrails for the standards of their particular codebase.
Bloated teams that diffuse ownership and responsibility.
Non-obvious tribal knowledge that was previously distributed through code reviews, but which now lack an AI-native distribution channel.
Amazon panicked.
They threw people at the problem:
Lukasz Olejnik@lukOlejnik
Amazon is holding a mandatory meeting about AI breaking its systems. The official framing is "part of normal business." The briefing note describes a trend of incidents with "high blast radius" caused by "Gen-AI assisted changes" for which "best practices and safeguards are not
6:35 AM · Mar 10, 2026 · 29.7M Views
978 Replies · 3.29K Reposts · 19K Likes
But people are fallible. The last thing a senior engineer wants to do is review reams of vibe-coded slop. And adding a human bottleneck eliminates any possibility of leverage from AI.
The way forward is defense-in-depth: as Ankit says, we’ll need multiple swiss-cheese layers of automatic guardrails: each may have holes (there’s no reason to believe 7.5 issues is anywhere close to the true number), but stacking differentiated review tools prevents problems from making it all the way through to production.
As an industry, we’ve been investing in the industrialization of software for over a decade: automated tests, CI/CD, deployment guards and automatic rollbacks, alerting and observability.
AI is a multiplier; companies accordingly need to multiply investment in tooling:
code review bots to catch bugs within PRs;
in-agent guardrails to stop problems before they get to PRs;
cross-team agent coordination to avoid changes that are independently reasonable but conflict with each other.
Every vendor would like you to believe that one more tool will make all your problems go away (and, full disclosure, we’re building @tanagram_ because we, too, want to make all your problems go away).
But the hard truth is that AI-native tooling needs to be paired with culture changes:
The previous generation of engineers were evaluated on their language and algorithmic proficiency. Coding well with agents requires high skill, and modern engineers should be evaluated on how well they can prompt and drive their agents.
Roles will blur; SWEs, PMs, PMMs, designers, support specialists, and more must all become product engineers. Everyone will need to talk to users and internal stakeholders, figure out problems worth solving, and implement maintainable solutions. This is not a prescription, but it will be an emergent phenomenon: your top performers are already blurring the lines, and everyone else must follow suit.
As roles blur, teams will shrink. Instead of running projects with a PM, designer, 5 engineers, and a legion of cross-functional managers that are impossible to get ahold of except when they block something, responsibility needs to concentrate into 2–3 people. Diffused responsibility leads to slop; concentration leads to ownership. Empower them to move fast with strong guardrails.
At the end of the day, enterprises will want to see ROI on all their AI spend.
Leaders may be tempted to cut headcount. I think this is the wrong approach.
Instead, I’d encourage more experimentation and shots-on-goal. Dig up ideas from your backlog. Let people build their ideas.
After all, Boris created Claude Code as a side project. It’s now on track to add an incremental Fortune 500’s worth of revenue to Anthropic.
The slopocalypse is here, but we’ll get through it. On the other side, the upside is unlimited.