Everyone is using it already. Now measure what it did
At 96% adoption the interesting questions are all downstream. Generation got cheap; review did not, and that asymmetry is where the returns actually get decided.
GitKraken's 2026 survey of 554 developers and leaders puts AI coding tool adoption at 96.4% of organisations, shipped alongside GitLens 19 and its reworked Commit Graph.
Adoption is a closed question
A number that high in a developer-tools survey should be read as "effectively everyone in this population." It is a self-selected sample and it does not matter — nobody seriously disputes the direction any more.
Which means the interesting questions all moved downstream, and almost none of them have published answers: what happened to defect rates, to review latency, to how long a junior takes to become useful, to how much of a codebase any one person understands.
The bottleneck moved and the tooling followed
GitLens consolidating coding, review, coordination and shipping into one surface is a response to volume. When generating code becomes cheap, everything downstream of code existing becomes the constraint.
Writing code was never the expensive part of software. Understanding, reviewing and maintaining it was.
A tool that accelerates the cheap step and floods the expensive one can raise output and lower throughput simultaneously. That is not a hypothetical failure mode; it is the default one, and it is invisible in any metric that counts lines or merged pull requests.
Why everyone keeps entering anyway
Meta launched a coding agent into this market this month, against incumbents, cloud vendors and IDE-native tools that were already there.
Nobody enters a market this crowded believing they are first. They enter because agentic coding is the clearest product-market fit AI has found — the one place where the model does a job that used to be a person's and the buyer can measure the result.
It is also a category with an honest quality signal, which cuts against late entrants. The code compiles or it does not. Where quality is legible, distribution advantages erode faster — and the environment they are launching into includes a model marketed on agent benchmarks at half price until January.
What to measure
Adoption is a lagging indicator of enthusiasm and tells you nothing. Track review latency, change failure rate, and the time from a defect being introduced to being found.
If those three moved the right way, the tools worked. If output rose and they moved the wrong way, you bought volume and called it productivity.
DevOps.com — Meta Launches AI Coding Agent to Challenge OpenAI and Anthropic → · Agentic.ai — Agentic AI News — August 2026 Launches, Models & Research →