CircleCI just dropped their 2026 State of Software Delivery report, and there’s a number in it that should make every founder and CTO pause: main branch success rates have dropped to 70.8%, the lowest in five-plus years. Nearly 3 in 10 production merge attempts are now failing.
This is happening at the exact same time that overall code throughput jumped 59% year-over-year.
Let that sink in. We’re writing dramatically more code and shipping it less reliably.
The speed trap
I’ve spent the last few months building with AI-assisted development tools daily. Claude Code, Cursor, the whole stack. The productivity gains are real. I can prototype working software in days that used to take weeks. I’ve been doing this for over 25 years and I’ve never seen anything like it.
But here’s what the CircleCI data confirms: writing code is no longer the constraint. Review, validation, integration, and recovery are where AI-generated code is piling up. The bottleneck has shifted, and most teams haven’t shifted with it.
The top 5% of teams saw a 97% throughput increase. The median team? Up just 4%. The gap isn’t about who has the best AI tools. It’s about who has the infrastructure, process, and judgment to actually ship what they build.
What the best teams do differently
After nearly a decade leading product and R&D teams, including managing a team that shipped supply chain software to the largest companies in the world, I’ve seen this pattern before. Every time a new force multiplier arrives (Agile, cloud, CI/CD, and now AI), the teams that win aren’t the ones that move fastest. They’re the ones that move fastest without breaking things.
That means:
Investing in the pipeline before the code. If your CI/CD isn’t rock-solid, AI-generated code will overwhelm it. The report shows recovery time climbed to 72 minutes, up 13% year-over-year. That’s not an AI problem. It’s an infrastructure problem.
Treating review as a first-class activity. When a human writes 50 lines of code, reviewing it is manageable. When AI generates 500 lines in the same time, you need a fundamentally different review strategy. I’ve found that having AI generate diff summaries and test plans alongside the code makes review tractable. Without that, you’re just rubber-stamping.
Keeping human judgment in the loop on architecture. AI is great at implementing. It’s terrible at deciding what to implement. The teams pulling away from the pack still have experienced engineers making the structural calls.
Why this matters for startups
If you’re an early-stage company and you think “we’ll move fast and figure out the process later,” the data says otherwise. Mid-sized companies (21-50 employees) had the worst recovery times in the report, nearly 3 hours per incident. That’s almost 4x slower than both smaller and larger companies.
The sweet spot is starting with good infrastructure habits from day one, using AI to accelerate the building, and keeping an experienced hand on the tiller. That’s the approach we take at Londonberry, and it’s why “rapid prototyping” doesn’t mean “reckless prototyping.”
Speed without infrastructure is just chaos with a faster clock.