5 takeaways from the State of Software Delivery Q3 Pulse report

Software teams are taking on more work than ever. Developers are juggling more branches, coding agents are keeping CI busy through the weekend, and builds are passing more often and finishing faster. But is all that extra activity helping teams ship faster?

Today we released the 2026 State of Software Delivery Q3 Pulse report, the latest in our quarterly series tracking how software teams are adapting to AI-driven development. Based on more than 22 million CircleCI workflows, it reveals how your team can be producing more code, validating it faster, and still losing ground to competitors who ship more.

Here are five takeaways worth your attention.

---

1. The velocity gap is now 10.4×

The highest-throughput software teams are pulling further ahead. The top 5% of main-branch workflows now run 10.4 times as often as the median, up from 9.1 times in Q2. The median workflow still runs 1.7 times a day, unchanged from Q2, while the top 5% are approaching 18.

The gains aren’t limited to the very top. Since Q2, throughput has barely changed for the bottom 75% of workflows. Above that, growth increases steadily, reaching 14% at the 95th percentile.

Teams that get changes onto main more frequently can ship features, respond to customer feedback, and improve their products faster. If you’re not increasing delivery velocity, you’re falling behind competitors who are.

But the widening gap doesn’t mean CI performance is getting worse. In fact, the Q3 data shows considerable improvement.

2. Delivery fundamentals are improving, but throughput isn’t keeping up

CI performance has improved considerably since the start of the year, across several important measures.

• More builds pass. Main-branch success rates climbed from 70.8% in Q1 to 78.3% in Q3.
• Workflows finish faster. Median main-branch duration fell from 2 minutes 25 seconds in Q2 to 2 minutes 2 seconds in Q3.
• Failures get fixed sooner. Median recovery time on main improved from about 59 minutes in Q1 to 54 minutes in Q3.

And teams aren’t improving their metrics simply by running fewer checks. Main-branch workflows now run 16% more jobs than in Q1, even as they finish faster and pass more often. Teams are running more validation steps per workflow without paying for it in longer build times.

Improvements in coding agents, developer experience with AI, and CI tooling may all be contributing. Together, they’re helping teams address the validation problems that were slowing delivery earlier this year. With less time spent waiting on builds and fixing failures, you’d expect more code to be reaching main.

Yet for most teams, delivery velocity hasn’t improved. Instead, the additional development capacity appears to be going somewhere else.

3. Teams are starting more work than they finish

Developers are taking on more work across more projects and branches, but they’re not getting changes onto main any faster.

Since Q1, the number of feature-branch workflows active on a typical day has grown 40%, compared with just 17% on main. Developers are juggling more parallel work, but individual workflows aren’t moving any faster. Feature-branch failures are even taking longer to resolve, with median recovery time increasing from about 80 to 86 minutes.

It’s easy to see how teams get here. When an agent can take on another task in seconds, why not give it one? But every change still needs review, validation, and integration before it can reach customers. As developers spread their time across more active branches and projects, unfinished changes and failed builds can sit longer before someone gets back to them.

The pattern extends beyond the workweek, too. Since Q1, feature-branch workflows active on a typical weekend day have grown 65%, compared with 21% on main. Agents and bots now account for roughly one in five weekend feature branches. Teams can keep more development work running around the clock, but they’re not increasing delivery at the same rate.

For most teams, AI is making it easier to start work than to finish it. The highest-performing teams, however, are managing to increase both development activity and delivery speed.

4. The top teams are taking on even more work. They’re shipping it, too.

The top 5% of teams by main-branch activity are working across 55% more feature branches than in Q1, compared with 30% growth for the median team. They’ve expanded their development activity even faster than everyone else, without a measurable increase in contributors.

But unlike most teams, they’re getting more of that work onto main. Since Q1, their main-branch throughput per workflow has increased 31%, and their total main-branch runs are up 57%.

Most teams are adding workstreams without increasing how frequently they deliver changes through each one. The leaders are doing both: expanding the amount of work they take on and increasing delivery velocity within each workstream.

For engineering organizations investing in AI, that’s the outcome worth pursuing: more software delivered without a corresponding increase in headcount. So what are the highest-performing teams doing differently?

5. The most active teams land changes in far fewer CI runs

One important difference is how many times teams run CI before changes reach main.

Merge efficiency ratio (MER) compares feature-branch workflow runs with main-branch runs. Including the final run on main, the median team needs roughly five CI cycles per shipped change. The top 5% need about 3.6, while the top 20 organizations need fewer than three.

Every additional cycle means another round of testing and potentially more debugging, fixes, and waiting. Across hundreds or thousands of changes, the extra time and compute costs add up. Teams that can get changes through validation in fewer cycles spend less effort on repeated work.

And despite improving CI success rates, the median team’s MER has barely changed since Q2. Builds are passing more often and finishing faster, but most teams still need roughly the same number of CI cycles to get changes merged.

Reducing unnecessary validation cycles is one way teams can make better use of their growing development capacity. Catch failures before the push, shorten feedback loops, give agents better tools to investigate and fix CI failures, and keep changes small enough to review and merge easily. Fewer rounds of testing and rework mean developers can spend more time moving changes forward.

---

The challenge ahead: Close the gap

AI is helping developers produce more code, and improvements in CI are making that code easier to validate. Yet most organizations aren’t shipping any faster. They’re spreading their capacity across more projects and branches without increasing how frequently changes reach main.

The highest-performing teams are showing what’s possible. They’re expanding their development efforts, getting changes through validation in fewer cycles, and increasing delivery throughput without adding developers.

As teams plan their next investments in AI and engineering productivity, the question is whether those investments will help developers finish more work or simply start more of it. Managing parallel development, reducing unnecessary rework, and bringing validation closer to the development process can help more teams turn increased development capacity into software that reaches customers.

Explore the full Q3 Pulse report to see where the leaders are pulling ahead and what your team can do to catch up.

Read the full Q3 Pulse report