Developers who used AI on at least 75% of their coding days merged a median of 95.7% more pull requests per week in May 2026 than they did in June 2025 according to LinearB's benchmarks of 2.7 million pull requests from 253 engineering organizations, yet at the top-performing organizations in that same sample only 79% of the pull requests opened by autonomous agents merge within 30 days against 92% for changes written by humans alone.
LinearB, a Leader in the first Gartner Magic Quadrant for Developer Productivity Insight Platforms, launched Agentic Studio to close that gap by giving engineering leaders and coding agents the same view of everything that has happened across software delivery. The launch matters to anyone paying for AI coding tools because the benchmark data shows the extra code piling up in review queues and the token bill climbing fastest at the top of the distribution.
The Launch: Engineering Context for Leaders and Agents
LinearB has spent years collecting data from the tools engineers use every day, such as GitHub and Jira, while linking each code change to the ticket that requested it and the release that shipped it. Agentic Studio puts that history to work for managers, who can type a plain question such as where delivery is stuck or whether AI spending is paying off and receive an answer with a suggested next step.
Coding agents get the same history through an MCP server, which is a standard way for AI tools to plug into outside data, along with add-ons that LinearB calls agent skills. An agent about to edit a file can now see who owns that code, how often it has been rewritten and whether it caused recent incidents, which tells the agent how carefully it needs to work.
https://x.com/LinearB_Inc/status/1724483081686655479?embedable=true
The first skill, called effort, is in early access and uses that history to decide how hard an agent should try on each change. A simple, low-risk edit gets a quick pass that uses fewer tokens, the units AI providers bill by, while a complex or risky change gets more effort with careful, tested edits. The skill works with any coding agent, Agentic Studio is available now to all LinearB customers and the next skills will focus on code review and merging.
AI Coding Spend
Companies are pouring money into AI coding tools, with enterprises spending $4.0 billion on them in 2025 against $550 million a year earlier according to Menlo Ventures. That sevenfold jump makes coding the largest single use of AI inside companies, at 55% of the $7.3 billion spent on departmental AI.
MarketsandMarkets expects the market for AI code assistants to grow from $8.14 billion in 2025 to $127.05 billion by 2032, a rise of about 48% a year. As writing code gets cheaper, the pressure moves to everything that happens after the code is written, which is why the tools that measure engineering work are growing fast too, with Gartner estimating that category at about $400 million and expanding more than 40% a year.
The Delivery Bottleneck: Why We Have Larger Pull Requests and Longer Waits
More code does not mean faster delivery, because every change still has to be checked by a human reviewer before it ships. LinearB's 2026 Software Engineering Benchmarks Report, which covers 8.1 million pull requests from 4,800 teams in 42 countries, shows that AI-assisted changes run to 408 lines of code against 157 for changes written without AI, measured at the 75th percentile, which makes them 2.6 times larger (our arithmetic).
Changes written by agents face a different problem, since they wait 17.6 hours for a reviewer to pick them up against 3.4 hours for human changes, which is 5.2 times longer (our arithmetic). Once a reviewer starts, the review itself takes 6.4 hours, so an agent's change spends almost three times as long waiting in the queue as it does being read. Across the whole dataset 84.4% of changes written without AI were accepted within 30 days against 32.7% of AI changes.
Pull request size and review wait by type.
Output: Heavy AI Users Pull Ahead
The productivity gains are real for heavy users, since among the 83,000 developers in LinearB's 2026 AI benchmarks those who used AI on at least three out of every four coding days merged a median 95.7% more pull requests per week in May 2026 than in June 2025. Developers who used no AI at all merged 3.6% fewer over the same period.
The gains grow step by step with usage, with developers using AI on half their coding days gaining 65.4% and occasional users gaining 24.4%. LinearB says these figures show a correlation that does not prove AI alone caused the gains. It also notes that the acceptance rate of heavy users' changes slipped 4.4% over the year. AI use is now close to universal, with 88.3% of organizations using AI tools daily or weekly, yet 44.7% still do not measure what those tools deliver.
Merged pull request growth by AI usage band.
Adoption Depth: What Elite Organizations Do Differently
Even the best teams use AI less than the hype suggests. At elite organizations, the top 10% in LinearB's sample, AI helps write only 54% of pull requests, AI writes 45% of the code that gets merged and fully autonomous agents open just 4.7% of all changes.
The biggest difference between the best teams and the rest is AI code review, where an AI tool checks a change before a human does. Elite organizations use it on 57% of pull requests against 8% at fair organizations, the top 60%, which is a sevenfold gap. It is also the easiest win on the table, because changes with an AI review are accepted at the highest rate of any type, lifting acceptance by up to five percentage points, while LinearB's August analysis found a lift of two to three points across almost every type of change.
Output Against Token Cost
More AI output comes with a much bigger bill, since organizations merge 2.6 pull requests per developer each week, twice the 1.3 at fair organizations, yet spend $481 per developer each month on AI tokens against $50, which is 9.6 times as much.
In plain terms, doubling output has cost close to ten times as much in AI usage, which is the trade-off that boards and finance teams now ask engineering leaders to explain. LinearB labels these token figures as rough averages, though the direction is clear: this is the gap the effort skill is designed to narrow by saving expensive AI effort for the changes that need it.
Agent Yield and the Ownership Gap
Acceptance is where agents still fall short, since at elite organizations 79% of the changes opened by autonomous agents are merged within 30 days against 92% of changes written only by people, a gap of 13 points (our arithmetic).
The gap widens at weaker organizations, where agent acceptance drops to 58% at good organizations and 37% at fair ones, leaving agent work 28 and 44 points behind the overall rate (our arithmetic). LinearB's explanation is ownership, because a change that no developer feels responsible for often never gets merged. That is why Agentic Studio gives agents code ownership history, so an agent knows whose code it is touching before it opens a change.
Cost Control: Token Spend and the Effort Skill
Token bills are rising fastest among the heaviest users, since at the 90th percentile, meaning the top tenth of spenders, developers now use $35.20 of AI tokens per coding day, up nearly 75% in three months, while the typical developer's spend has barely moved. That works out to $481 a month or $5,772 a year per developer (our arithmetic), which LinearB says is still under 4% of what a developer costs in total.
Gartner expects spending on AI coding tools to pass the average developer's salary by 2028 as token use keeps climbing. LinearB's own worked example, which is an illustration and not customer data, shows why the cost of each delivered change matters more than the raw bill, since a 350-developer company that quadruples its token spend from $300,000 to $1.2 million a month while lifting output from 3.3 to 5.67 merged changes per developer each week cuts its cost per delivered change by 29%, from $864 to $613.
Customer Evidence: Kraken and Emburse
Kraken offers an early example of the data at work, since the company used LinearB's MCP server with an AI assistant to find out what was slowing its code reviews and learned that the cause was time zones and reviewer availability, not large changes. Over the next four months it cut its slowest review times, measured at the 90th percentile, by more than half.
Emburse uses the data at board level, with chief technology officer Ken Ringdahl putting LinearB charts directly into board decks because his board wants to know how the company uses AI and what return it gets. LinearB says its AI assistant and MCP access have the highest user retention of any features it has released.
What Comes Next
LinearB plans its next agent skills for code review and merging, followed by tools that help teams govern AI-written code safely and cut what they spend on AI. Chief executive and co-founder Ori Keren describes the company's data as a knowledge graph of the entire software delivery process that coding tool makers do not have, while co-founder and chief operating officer Dan Lines says the goal is to move leaders from reading dashboards to knowing what to do next.
For engineering leaders, LinearB's data points to three places where AI coding is losing value today, which are the review queue, the agent change that nobody owns and the top of the token bill. Agentic Studio is LinearB's attempt to fix all three by giving people and agents the same context, in a market that Gartner sizes at about $400 million and growing more than 40% a year.
Don’t forget to like and share the story!
Vested Interest Disclosure: HackerNoon has reviewed the report for quality, but the claims herein belong to the author. #DYOR.
