DX Q3 2026 and the cost of trusting code (Series #AI4SDLC)
In two episodes of Research Insights Made Simple, Evgeny Sergeev and I looked at how DX proposes measuring engineering productivity and the impact of AI. Here are the podcasts on DX Core 4 in general and measuring AI impact. DX applies both frameworks in its October report. Core 4 covers delivery speed, developer experience, quality and the share of time spent on new capabilities. The AI Measurement Framework adds AI usage, impact and spend.
The report draws on telemetry and surveys from more than 500 organizations using DX. From Q4 2025 to Q3 2026, median weekly TrueThroughput rose 75.5%, from 1.43 to 2.51 weighted PRs per engineer. In the last quarter developers gave noticeably better ratings to production debugging (+13.3%), documentation, and build and test (both +10.4%). The score for ease of shipping changes rose 3%. The change failure rate and recovery time after a failed deployment stayed roughly flat.
In surveys, time spent on new capabilities moved from 57% to 59% over four quarters. Telemetry put about 56% of observed AI sessions in maintenance and fixes and 11% in new capabilities. I like having two ways to look at the team’s work. The survey tells us how engineers divide their time, while sessions show what they call an agent in for. To find out where saved hours went, you’d have to follow the tasks from start to finish.
Now for the notes I made in the margins.
1️⃣ What went into that 75.5% increase? TrueThroughput uses a model to weight merged PRs by complexity. Diff size in lines is one input, and the median PR grew 122% in that period. I’d like to see plain merges per engineer and average PR weight side by side for the same companies. The chart makes it hard to tell how much growth came from extra merges and how much from changes DX scored as heavier. Page 10 also mentions a 47% increase in delivery volume. I’d like to ask DX what they counted there. Abi Noda touched on this in the DX newsletter when discussing the model weights.
2️⃣ Why the longer wait for first review? In Q3, companies with 20–40% AI code had a median wait of 40 minutes, versus 82 minutes in the group with 80% or more. Alongside wait time I’d look at PRs per reviewer and active review time per PR. If reviewers are busy with other changes, you can redistribute the load. If they spend ages reconstructing context for each change, I’d look at what gets submitted for review.
3️⃣ AI spend per developer rose roughly 21-fold in a year I circled both points on the chart: $63 in Q4 2025 and $1,314 in Q3 2026. Cache reads accounted for 94% of tokens in the final quarter. To tell how much money goes on rereading context, we need billing broken down by token type. Next I’d calculate the cost of one task that passes review and tests and see how that price changes as spending rises. This picks up my post on token burn.
4️⃣ The cost of trusting a change DX suggests that some review work goes into recovering what was missing from the agent’s context. Why is this change needed? What constraints apply? How will we know the task is done? The authors propose evaluating context with the CAFE(S) framework. In Q3 this is still a hypothesis about the causes of delays. I’d like to test it on comparable tasks. If we write down architectural constraints and acceptance criteria in advance, how much human time can we save from opening a PR to merging it, including rework?
#AI #AI4SDLC #Engineering #Metrics #Research
Files from the post
- Annotated_DX_2026_Q3_Report_State_of_AI_Impact_in_Engineering.pdfDownload PDF