Compared with a year ago, the same engineers in Sidetrade R&D open 124% more pull requests a week and review 125% more of their colleagues' work.
Numbers like that deserve suspicion, including mine. So, here is a transparent assessment of what we measured, what it covers, what got worse, and how far we still have to go.
The operating model behind the numbers
I'd point first to two major changes over the year. Firstly, every engineer got AI coding tools, and secondly, from March 2026 the teams moved onto PodFlow, the AI-native R&D operating model Jamie Hennerley described in September. PodFlow turns every stage of delivery, from discovery to release, into a versioned skill that runs inside the tools our engineers already use. A person owns every transition between stages. By June, every R&D team was running it.
I cannot split the gain cleanly between the tools and the operating model, and I will not pretend to. What I can show is where the curve bent. Across the full roster, pull requests per engineer grew 17%, 6%, and 19% in the three quarters before every team was on PodFlow. In the first quarter with every team on it, they grew 39%.
Twelve months of delivery data
The headline figure is 2.1×, the average of the gains in pull requests, reviews, and commits for the engineers who were here in both Q3 2025 and Q3 2026. Measuring the same people removes the easiest way to flatter a number like this, which is changing who gets counted.
Output, same engineers
- Pull requests per engineer: +124%
- Reviews per engineer: +125%
- Commits per engineer: +70%
- Lines changed per engineer: +321%
Pace, all engineers
- Story points per working day: +80%
- Engineers at or above 40% of reference pace: 35% to 70%
- Median pull request cycle time: -79%
Quality checks, all engineers
- Test code: 32% to 47% of changed lines
- Merge rate: 87% to 89%
- Approved before merge: 79% to 75%
Each figure shows the change from Q3 2025 to Q3 2026.
The average of the pull request, review and commit gains.
A like-for-like comparison, with the same engineers measured in Q3 2025 and Q3 2026.
Every R&D engineer in each quarter.
What these numbers cover
Pull requests, reviews, and commits measure activity. The standard objection is that AI tooling inflates them by slicing work into smaller pieces, so I checked. The median merged pull request grew 220% over the same year, and lines changed per engineer rose 321%, close to three times the growth in pull request count. The work got bigger as well as more frequent.
The guardrails mostly held. Tests went from 32% to 47% of changed lines. The share of pull requests that merge rather than get abandoned stayed between 87% and 89% all year, and median cycle time from open to merge fell 79%.
None of this is a customer outcome. It tells me engineering capacity roughly doubled. Whether that capacity reached customers as more value is a separate measure, and I will report it when I can defend it.
What got worse
Large pull requests, over 500 changed lines, rose from 24% to 40% of the total, and the size of the largest tenth roughly tripled. Over the same year, the share approved by a reviewer before merge slipped from 79% to 75%.
More code, in bigger changes, with a slightly smaller share of it approved by a second person. Nobody decided to make that trade. Generating code got cheap and reading it did not, so review has become the bottleneck, and the next phase of our operating model has to be designed around it.
The largest gain left is adoption
Every engineer has the same tools and the same operating model. A year in, 70% of them work at or above 40% of the reference pace, up from 35%. That still leaves nearly one in three below it.
Installing a tool takes an afternoon. Changing how an experienced engineer plans, writes, and reviews code takes months, and a process document has never made it happen on its own. What moved our numbers was engineers seeing a colleague on the same codebase work differently and deciding to try it. It is slow, it varies from one person to the next, and I have not found a way to rush it. In the end, this is a change management job, and that is where our next big gains will come from.
The third of the team still finding its way in is also the largest single source of gain left in the data. Closing that gap is a management job first and a tooling job second.
Three sources for the rest of 3×
Our O2C Intelligence 2030 strategic plan commits R&D to 3×. At 2.1×, we are a little past halfway. Most of the first half came from making engineers faster at the coding stage. The second half comes from changing the shape of the work itself. That means agents carrying tasks end to end, and a software factory approach to the whole path from idea to release.
Today an agent drafts, a person reviews and a person merges. The next step is agents taking a task from ticket to tested pull request inside the gates PodFlow already defines, with a person still owning every transition.
The next version of the operating model is built around the bottleneck this data exposed. A change should reach its reviewer small enough to read and carrying the evidence that it works, so approval before merge goes back up while cycle time keeps falling.
Once code moves from open to merge this fast, the stages around it set the pace. One is turning an idea into a story an engineer can start. The other is getting merged code tested and released to customers. That is where we are looking next.
Questions this post answers
How much did developer output change at Sidetrade with AI coding tools and PodFlow? For engineers present in both Q3 2025 and Q3 2026, pull requests per engineer rose 124%, code reviews 125%, and commits 70%. Sidetrade reports this as a 2.1× overall delivery impact, against a 3× target set in O2C Intelligence 2030.
Did pull requests get smaller as AI coding tools were adopted? At Sidetrade, they got larger. The median merged pull request grew 220% over the year, and lines changed per engineer rose 321%, faster than pull request count.
What is PodFlow? PodFlow is Sidetrade's AI-native R&D operating model. It packages the delivery process from discovery to release as versioned skills, with a person owning every transition between stages, and every R&D team was running it by June 2026.
What limits further AI productivity gains in software engineering? At Sidetrade, review and people adoption. Large pull requests rose from 24% to 40% of the total while the share approved before merge fell from 79% to 75%, and nearly one engineer in three still works below 40% of the reference pace despite having the same tools as everyone else.
