From Pull Panda to DX Core 4
GitHub acquired Abi Noda’s Pull Panda tools for reviewing and analysing pull requests in 2019. The data showed where a change was waiting and led to a broader question about what developers found difficult in their daily work. DX launched in 2022 with a focus on gathering those observations. The 2023 DevEx paper described three conditions that shape the experience: feedback loops, cognitive load, and flow. DX released Core 4 in late 2024. Atlassian completed its acquisition the following November.
Each of Core 4’s predecessors asks a different question. DORA looks at the pace and stability of delivery. SPACE brings outcomes, collaboration and well-being into view. DevEx explains why a team can have good delivery numbers and still struggle to get work done. Core 4 puts those perspectives into a shared conversation between teams and their leaders.
What the four dimensions measure
For speed, DX counts merged pull requests per engineer. For effectiveness, the Developer Experience Index, or DXI, captures survey responses about factors such as deep work, local iteration and the release process. Change failures and recovery time provide a view of quality. Business impact starts with a question about where engineering effort goes. The innovation ratio measures the share of time devoted to new capabilities. Progress and returns on individual initiatives call for their own measures.
Surveys and telemetry illuminate different parts of the work. Engineers can report what slows them down and how much time they spend on new capabilities. Repository logs can record merged changes. The share of PRs classified as bug fixes is a different proportion from the share of production changes that fail. A target based only on PR count can reward splitting work. I would put any increase in throughput alongside quality and developer experience before drawing a conclusion.
AI use, outcomes and the price of a result
I used the AI Measurement Framework in fintech to agree on three questions. Who uses the tool and for what? What changes in task time, throughput or quality? What does it cost to get that result? Active users and the share of AI-authored code describe uptake. Time saved, pull request throughput, failures and confidence in releases describe different consequences. Licences and tokens are easy to find on a bill. Review, rework and cost per successful task need their own accounting.
In the episode I work through a hypothetical DX calculation for 200 engineers. DX values the extra throughput as equivalent engineering capacity. With an estimated benefit of $3.2 million and costs of $0.8 million, the model yields a net benefit of $2.4 million and a 300% return. The result depends on whether gains persist, what costs are included and whether reliability holds up. An AI-assisted dependency migration might pay for itself. Scaling the same estimate to an organization changes the work mix and the costs.
Reviewing the context an agent receives
CAFE(S) looks at five properties of context. Clarity asks if the instruction makes sense. Actionability asks if the agent can take a next step. Fidelity concerns accuracy and currency. Efficiency concerns how readily the agent can find what matters. Security covers access to information and permission to act. I apply those questions to an instruction in AGENTS.md. Specify the working directory, the test command, the environment and what success looks like. Check the guidance against CI and give the agent the access it needs for that task. The example illustrates a way to review context and leaves the size of any productivity gain open.
Reading the 2024–2026 reports
In the 2024 DevEx survey, 69% of developers reported losing at least eight hours a week to friction. In DX’s Q4 2025 report, developers estimated that AI assistants saved them 3.6 hours a week on average. One is a share of people above a threshold for time lost. The other is a mean estimate of time saved with a tool. Saving time writing tests does not show whether a team is spending any less time waiting for access to an environment.
In a sample of more than 400 companies with rising adoption, AI tool use grew by 65% on average from November 2024 to February 2026. Median growth in unweighted PR throughput was 7.76%. The Q3 2026 report shows median weekly TrueThroughput rising from 1.43 to 2.51 complexity-weighted PRs per developer from Q4 2025, up 75.5%. The first result summarizes company-level changes. The second compares quarterly levels of a weighted metric. Historical figures can also change when additional customers bring historical data into the sample.
Across the quarterly sample, aggregate AI spending rose from $10 million to $425 million. That total does not describe a typical company. Developers reported more noticeable gains in local tasks, including production debugging and documentation, than in direction or maintainability. Median time to first review was 40 minutes among organizations with a 20–40% AI code share and 82 minutes among those at 80% or above. The groups may differ in PR size, reviewer availability and review processes. The comparison gives teams a reason to inspect their own queues and look for where the time saved in drafting changes went.
What to take away
- 01Read faster throughput alongside quality, working conditions and how engineering effort is allocated.
- 02Measure adoption, outcomes and the cost of accepted work separately. Include review and rework when assessing return.
- 03Before comparing reports, check the metric definition, unit, denominator, window, sample and aggregation method.
Sources
- Episode slides
- Podster audio recording
- Episode 12. Measuring Developer Productivity With DX Core 4 — mentioned at ≈14:10 (Podster)
- Episode 30. Developer Productivity: Google’s Human-Centered View — mentioned at ≈17:05 (Podster)
- Episode 11. Measuring AI Code Assistants and Agents — mentioned at ≈20:04 (Podster)