AI Engineering Trends

Jellyfish measures Autonomous Agent Activity as the percentage of pull requests that are autonomously created or generated by AI agents. A PR qualifies as agent-generated if it was opened by a user identity that is an agent, or if it contains commits where the committing user is an agent. This metric tracks the emerging frontier of AI adoption, where AI isn’t just assisting engineers, but independently shipping work.

Agent involvement took its largest single-month step in September: the median company now has 28% of its merged PRs with agent involvement, and leading companies 58%.

Jellyfish measures Software Factory PRs as the share of a company’s merged pull requests that were opened by a software factory agent: an autonomous system that takes work from specification to pull request without an engineer driving the tool. It is calculated monthly per company, shown as the top decile of companies (P90, P95, and P99), because fewer than one company in ten merges any factory-opened PR.

Jellyfish measures Model Usage as the share of each tool’s active users who used a given model, reported separately for Claude Code, GitHub Copilot, and Cursor. It is calculated from the models invoked by each developer each week. Users can appear under multiple models. 

By analyzing the top three models by share of users each month, we measure how quickly engineering organizations absorb new model generations. As each new release ramps, its predecessor fades, so the top models can change from month to month.

Jellyfish measures open weight model usage two ways: the fraction of engineers who selected an open-weight model (e.g. Kimi, GLM, DeepSeek, Qwen) in a given week, and the fraction of all model tokens that went to open-weight models. Both are calculated from per-event model selections in coding tools, classified by model family, and describe developer usage rather than production usage.

The two lines tell the story of a niche that many tried and few kept. Fewer engineers chose an open-weight model in September (0.5% in the last week, down from the 0.7% August peak but still about 3.5x the March level), while open-weight models’ share of all tokens held at its high of 0.3%: those who stayed use more, and the volume is concentrated in fewer companies.

Jellyfish measures open weight model usage as the fraction of engineers who selected an open-weight model (e.g. Kimi, GLM, DeepSeek, Qwen) in a given week, alongside which model families lead among the engineers using them and where the open-weight token volume goes. It is calculated from per-event model selections in coding tools, classified by model family, and describes developer usage rather than production usage.

As software organizations mature in their AI transformation, their AI budgets are growing correspondingly. Teams are employing a number of strategies to manage that spend, including leveraging open-weight models and routing usage through model routers and gateways. This behavior remains concentrated in the most AI-mature companies, with under 2% of companies employing open-weight models or routers regularly. In September 0.5% of weekly engineers used an open-weight model, down from the 0.7% August peak but still about 3.5x the March level.

The two rankings split “most adopted” from “most used.” Kimi (Moonshot) leads adoption at 63% of open-model engineers, while GLM (Zhipu) leads volume at 53% of open-weight tokens to Kimi’s 33%. Qwen and DeepSeek follow on both rankings, and the four together carry 95% of open-weight token volume.

Jellyfish measures agentic-merged PRs as the fraction of a company’s merged pull requests whose only review activity, whether comments or formal reviews, came from AI tools, with no human review input before merge. It is calculated per company over merged PRs opened by humans or AI agents. The data below shows the distribution across companies by percentile.

This metric captures the sharpest edge of AI adoption: work that ships with AI as the only reviewer. While the median company auto-merges just 0.5% of its PRs, lead adopters of this workflow merge 9% of their PRs with no human review, and the gap between the two keeps widening.

Jellyfish measures invocations of agent skills in Cursor and categorizes them by use-case using an LLM. The data above shows the distribution of skill categories invoked by engineers within Cursor in the past month. 

This captures the current state of AI development practices and the specific workflow steps engineers are augmenting within the AIDLC.

Jellyfish measures AI Spend as the cost of AI coding tool usage per developer per week, across Claude Code, Cursor, Codex, and GitHub Copilot. It is calculated by summing each developer’s reported usage costs each week, shown as a distribution across developers.

This metric turns token consumption into budget terms, and the distribution matters as much as the level: spend is highly concentrated, with the heaviest users consuming an order of magnitude more than the median developer. For leaders, it anchors the ROI question by showing what the organization actually pays per developer for the productivity signals measured alongside it.

We show the relationship between AI Spend and Headcount as one point per company: how much its AI spend per engineer changed from the first quarter of 2026 to the third, against how much its engineering headcount changed over the same period. It is calculated for companies with complete billing data across the period, with measurement artifacts screened out.

The fitted line is flat, meaning AI spend and headcount appear to move independently: we are not seeing widespread, systematic changes in headcount as AI spend increases. The median company ended the third quarter at 98% of its first-quarter headcount while spending 3.8x as much per engineer.

Jellyfish measures Token Consumption as the total number of tokens consumed by a developer while using Claude Code, Cursor, Codex, and GitHub Copilot CLI. It is calculated by summing all tokens as reported from each tool for each developer each week. This metric tracks usage of those tools and the potential budgetary implications of increased AI maturity. The median developer consumed 59M tokens per week in September, about 8x the level of a year earlier.

Jellyfish measures Productivity by Token Consumption at the developer-week level. Each developer-week is ranked by its AI token consumption (Cursor + Claude Code) and grouped into 50 equal-sized bins, each 2 percentiles wide. Each point is one bin’s average tokens against its average merged PRs, and the line is a curve fitted to those points.

This metric addresses the core ROI question of whether consumption converts into output. Throughput rises with token consumption, but not linearly: gains taper off at the high end. Increasing the tokens used in a week tenfold, from the median developer-week’s level, buys about 2.2x the merged PRs. As with all comparisons in this report, the relationship is correlational rather than causal.

Jellyfish measures Cost per PR as the all-in cost of a merged pull request: engineering salary cost plus priced AI usage, divided by the number of human-opened PRs merged, computed for each company and shown as a distribution across companies. It is calculated monthly over the trailing twelve months.

This metric puts AI investment and delivery on the same scale. Over the past year the median company’s cost per merged PR fell by about a third, from $1,202 to $820, and the spread between the cheapest and most expensive companies narrowed. The driver is throughput. Companies merged more PRs per engineer while salary cost held about flat, and AI’s direct slice of the cost is small, about $18 of the $799 average.

Jellyfish measures Access Percentage as the fraction of engineers at a company who have a license to an AI coding tool. As a baseline adoption metric, Access Percentage establishes the foundation for deeper analysis, including how frequently and deeply AI tools are integrated into developer workflows.

Jellyfish measures Weekly Active Users (WAU) percentage as the fraction of engineers at a company who actively use an AI coding tool in a given week. This metric captures the frequency of adoption, showing whether engineers are actually integrating AI tools into their regular workflows.

Jellyfish measures AI Code Percentage as the fraction of a company’s shipped code that is AI-assisted. It is calculated as the fraction of merged code additions that were AI-assisted, relative to all code additions in merged pull requests for each company. 

This metric moves beyond tool usage frequency to capture the actual depth of AI’s impact on codebases and coding work.