AI Made Coding Faster. It Didn’t Make Shipping Faster.

We are getting very good at making code cheap.

A developer with Cursor, Claude Code, Codex, or another coding agent can now produce changes at a rate that would have looked ridiculous a few years ago.

But writing code was never the whole software delivery process.

Someone still has to understand the change. Review it. Run the tests. Investigate failures. Check whether it breaks something outside the diff. Merge it. Deploy it. And eventually figure out whether it actually worked in production.

This creates a weird situation: a developer might become 5x or 10x faster at producing changes while the organization becomes only slightly faster at shipping them.

There is already some evidence of this.

CircleCI’s 2026 software delivery data found that, for the median team, feature branch throughput increased by 15%, while main branch throughput fell by 7%. More code was being produced upstream, but less was making it through the delivery system.

DORA has observed a similar effect. AI tends to amplify the system around it. It speeds up code generation, but some of that saved time is simply moved into auditing and verification. Higher AI adoption has been associated with both higher throughput and higher delivery instability.

Even Cursor ran into this internally. As its coding agents became more capable, the team found itself spending more time reviewing their output. That problem eventually led to Bugbot, its automated code review agent.

This feels like the next obvious bottleneck.

If agents can open 20 pull requests in the time developers used to open four, asking humans to review those 20 pull requests in the same way does not scale.

The interesting problem is no longer just:

How do we generate more code?

It is:

How do we decide what is safe to ship?

That probably means much more automation around validation: testing changes in isolated environments, understanding their blast radius, checking security and API contracts, detecting risky changes, automatically fixing CI failures, and routing human attention only to the parts that actually need it.

Atlassian describes this as an orchestration problem rather than a prompt engineering problem: faster individuals do not automatically produce faster engineering organizations because work still gets stuck in queues, handoffs, and approvals.

That distinction will matter more as coding agents improve.

If AI makes code generation 10x faster but production throughput only improves 2x, the biggest opportunity may not be another coding agent.

It may be everything between “the code was written” and “the change is safely running in production.”

References

• CircleCI — Five takeaways from the 2026 State of Software Delivery Reporthttps://circleci.com/blog/five-takeaways-2026-software-delivery-report/
• Google DORA — State of AI-assisted Software Development 2025https://dora.dev/research/2025/dora-report/
• Cursor — Building a better Bugbothttps://prod.cursor.com/blog/building-bugbot
• Atlassian — From prompts to orchestration: Scale AI coding agent impacthttps://www.atlassian.com/blog/development/scale-agent-impact-with-jira-automation