AI Coding Agent Adoption: The 2026 Numbers
A report covering 110 million pull requests says agents touch 28% of merged code but only 0.5% lands unreviewed. The second number is the one that describes how teams really work.
The most interesting numbers on AI coding agent adoption published this week are not the ones being quoted. Jellyfish released its AI Engineering Trends report on 27 October 2026, drawing on 110 million pull requests from 294,000 engineers at more than 1,300 companies. The figure that travelled fastest was a per-tool league table of which agent merges the most code. That figure is not on the report's own page.
Start with what the report does say, because it is more useful than the league table would have been.
The numbers the report publishes
At the median company, 28% of merged pull requests had agent involvement in September 2026. At leading companies, 58%.
Automated merges with no human review: 0.5% at the median company, 9% among lead adopters.
The median developer consumed 59 million tokens per week in September, around eight times the figure a year earlier.
A tenfold increase in weekly tokens from the median developer-week is associated with about 2.2 times the merged pull requests. The report calls this relationship "correlational rather than causal".
Cost per merged pull request at the median company fell from $1,202 to $820 over the past year. The average is $799, of which AI's direct share is roughly $18.
The median company ended Q3 at 98% of its Q1 2026 headcount, while spending 3.8 times as much per engineer on AI. The fitted line between AI spend and headcount is flat.
Open-weight models remain marginal: 0.5% of weekly engineers used one in the last week of September, and under 2% of companies use open-weight models or routers regularly.
Definitions matter for the headline one. A pull request counts as agent-involved if it was opened by an agent identity or contains commits from an agent. That is a broad bar. One agent-authored commit in a human-driven branch counts.
The figure to be careful with
Secondary coverage of the report, including this write-up, reported per-tool shares of merged pull requests, naming one tool at 31% and another at 23%, from a sample of 100 million pull requests and 276,000 developers.
Two things do not line up. The sample on the report's page is 110 million pull requests and 294,000 engineers, which suggests the smaller numbers come from an earlier edition. More importantly, the report's page does not publish per-tool shares of merged pull requests at all. What it publishes is per-tool model usage, meaning the share of each tool's active users who invoked a given model, and it notes users can appear under more than one model. That is a different quantity with a different denominator.
None of this means the 31% figure is invented. It may sit in a section not on the public page. But if you are about to put a per-tool merge share in a slide deck, the primary page is not where it comes from, and the number you can actually cite from that page is a different one.
What AI coding agent adoption really looks like
The 0.5% line is the one to sit with. Agent involvement at 28% of merged pull requests sounds like a workforce being replaced. Unreviewed automated merges at 0.5% says something almost opposite: in practice, humans are still the gate on essentially every change that ships.
Put the two together and the shape of real adoption appears. Agents are writing a large and growing share of the code. People are still deciding what lands. Even among lead adopters, 9% unreviewed means more than nine in ten changes get a human look. The report's note that fewer than one company in ten merges any factory-opened pull request points the same way.
That is a workflow pattern, not a transitional phase, and it explains the cost numbers. Cost per merged pull request fell by about a third, but AI's direct share of that cost is around $18 against an average of $799. The money is still mostly people. What changed is how much each person gets through, which is also why headcount at the median company sat flat at 98% rather than falling.
What to take from this if you run a small team
Three practical readings:
Measure involvement and review separately. A dashboard that only counts agent-authored commits will tell you adoption is high while saying nothing about whether anyone is checking the output.
Treat the token-to-throughput relationship with suspicion. The report says it outright: correlational, not causal. Teams that already ship a lot may simply also use more tokens.
Do not budget as though AI spend replaces salary. The data shows spend up 3.8 times per engineer with headcount essentially unchanged.
If you are deciding how much autonomy to grant, we have gone through the tradeoff directly in should you let AI agents merge their own pull requests, and the practical checks belong in reviewing AI-generated code before you ship it. For the underlying definitions, see what an AI coding agent is, and for the wider tool landscape, our overview of AI coding tools.
Frequently asked questions
What share of merged pull requests involve an AI agent?
At the median company in the dataset, 28% in September 2026, rising to 58% at leading companies. A pull request counts if an agent opened it or contributed any commit to it.
Do AI agents merge code without human review?
Rarely. The median company reports 0.5% of pull requests merged automatically without human review, and 9% among the heaviest adopters.
Has AI coding spend reduced engineering headcount?
Not in this data. The median company ended Q3 2026 at 98% of its Q1 headcount while spending 3.8 times as much per engineer on AI, and the report describes the fitted relationship between spend and headcount as flat.
Which AI coding tool merges the most pull requests?
The report's public page does not publish that ranking. Figures circulating in coverage attribute per-tool merge shares to the report, but the page itself reports per-tool model usage instead, which measures something different.
How did this land?
About the author

Staff Engineer, Platform
Carlo works on the platform that turns prompts into running apps. He writes the engineering deep dives and the changelog notes worth reading.


