AI coding tools tripled deploy volume. Your CI/CD wasn’t built for that.
A platform engineering lead at a mid-sized fintech watched his team’s deployment volume triple over about a year. No new hires. No longer hours. Half the team was writing code with Claude and Cursor, and the other half was doing the same without mentioning it in standup. The output just kept climbing.
That’s the situation a lot of engineering orgs are quietly in right now, and most of their CI/CD pipelines were not built for it.
The assumption that broke
For the past decade, CI/CD infrastructure was designed around one core assumption: a human sits down, writes code over an hour or a day, and pushes it. That assumption is no longer true for a growing share of commits. AI coding assistants didn’t make engineers write better code. They made engineers write more code, faster, with far less friction between having an idea and having something committable.
The numbers reflect it. OpenAI reportedly runs around a million builds a day internally. A startup called Hermetic, barely three weeks old when it was mentioned publicly, was already handling half a million of those builds. Build volume is exploding because the bottleneck of writing code has been partially removed.
What breaks first
Build and test infrastructure cracks earliest. Queue times that used to be two minutes stretch to fifteen. Engineers get impatient and start skipping steps rather than waiting, which is when shortcuts compound into real problems.
Audit trails come apart next, and that failure is sneakier. When a human writes and commits code, there’s a clear story of who did what and why. When an AI agent generates and pushes changes on an engineer’s behalf, that story gets murky. Six months later, when you’re reconstructing why a specific change went to production, that ambiguity costs someone a very bad afternoon.
Cloud costs are the third pressure point, and they tend to hit finance before they hit engineering. Ten times the deployment frequency doesn’t automatically mean ten times the value delivered. It can mean ten times the compute burned on iterations that never should have made it past a first draft.
Three lanes, not one
The structural fix the author proposes is separating the pipeline into three lanes rather than scaling a single one:
- Fast lane: lightweight, disposable environments for genuine experimentation, minimal gatekeeping, nothing real at stake yet
- Staging lane: fuller test suite runs once something looks promising and is being considered for a real merge
- Production lane: every existing gate, plus a hard requirement that a human signs off before anything crosses in, regardless of whether a person or an agent wrote the change
The piece that ties it together is a unified audit trail across all three. Every action needs an identity attached: who or what triggered the build, what model or tool generated the change, who reviewed it before it moved lanes. Role-based access should scope agent permissions tightly to the lane they’re operating in. An agent in the fast lane should not be able to reach production just because someone granted it broad access once and forgot to revisit it.
The review layer hasn’t caught up
The deeper issue isn’t that AI tools write bad code. It’s that the review and governance layer around them hasn’t caught up to how much trust teams have quietly started extending to their output. An engineer who ships ten times more code and reviews every single change carefully is in a different position than one who ships ten times more and trusts the tool got it right. Most teams haven’t consciously chosen one path. It happened by default, shaped by habits that were in place before the tools got fast enough to matter.
The pipeline structure that felt safe six months ago may already be too permissive today, simply because the agents operating inside it have gotten more capable. The author’s recommendation: treat permissions, lane definitions, and audit setups as a regular cadence, not a one-time configuration task.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
