Coding agents have moved from novelty to routine in a matter of months, but full automation of software delivery remains rare. That is the central finding of Jellyfish's October AI Engineering Trends report, released Oct. 27 and covered by R&D World. The Boston-based engineering-management company analyzed activity data from more than 1,300 companies, 276,000 developers and 100 million pull requests.

Adoption is broad, autonomy is narrow

In September, AI agents opened or committed code to 27.7% of pull requests at the median company, compared with roughly 2% in January. At the 90th percentile, agent involvement reached 57.9%. The median share of code written by AI tools was 29%. Despite that growth, humans remain in the approval loop: the median share of merged pull requests approved on AI review alone was just 0.5%, rising to 8.5% at the 90th percentile, up from about 4% at the start of the year.

Software factories, defined as pipelines where agents choose, write, review and ship work with minimal human involvement, are still confined to a small group of early adopters. Their share of pull requests rounded to 0.0% at the 90th-percentile company, and only at the 99th percentile did factories open 9.4% of pull requests, up from about 1.7% in March.

Cost of agentic coding

Spending on Claude Code, Anthropic's coding tool, hit a median of $63 per active developer per week by late September. Using the report's four-week monthly conversion, that equals $252 a month, while the 90th-percentile monthly equivalent reached $1,666. These figures show that agentic coding is becoming a meaningful line item rather than a cheap productivity add-on, and that usage intensity varies enormously between organizations.

What a software factory requires

Nikolas Albarran, Jellyfish's AI product lead, describes software development as four stages: planning and work definition, building, review and shipping. A factory automates each with little human input. In most current pipelines, people still decide what work to do and define the goal. The most mature setups have an agent or orchestrator creating tickets and handing them to other agents for execution, which Albarran treats as a sign that the organization trusts agents with context and infrastructure. The idea has a long lineage; Hitachi launched its Software Works in 1969 to apply manufacturing principles to software.

Nicholas Arcolano, Jellyfish's head of research, argues the payoff is scaling beyond human attention. If a person must supervise every step, throughput hits a ceiling. His advice is to hand agents complete tasks, even simple and safe ones, rather than fragments of work.

Inflection point and context engineering

Arcolano places a turning point around the end of 2025, when agents became reliable enough for substantial work. Andrej Karpathy voiced a similar view in a February post on X, saying coding agents largely did not work before December and largely have since, and describing an agent that built a home-camera dashboard on its own in about half an hour.

The report also quantifies the value of maintaining context files, the instructions agents read about a project. Engineers who edited them six or more times merged 2.3 more pull requests per week than those who never did, controlling for AI usage, and each doubling of context-file investment was associated with 29% more merged pull requests. In Cursor, code review accounted for 14.9% of classified agent-skill invocations in September, followed by planning and tickets at 11.9%, suggesting developers are already applying agents beyond raw code generation.

The governance gap

A survey of 1,417 developers at 30 companies, run from January to September, found 88% agreed AI coding tools helped their work, but only 62% agreed their organization manages the risks of AI-generated code effectively. Arcolano said some companies are making progress in pockets, but very few have figured out how to scale it to everyone, and that training and change management have lagged the pace of tool improvement.

Analysis

The data points to a gap between capability and organizational readiness. Agents are touching a large share of pull requests, yet the bottleneck for deeper autonomy is trust: reliable context, review controls, and trained developers. For vendors such as Anthropic and Cursor, high per-developer spend signals real demand, while the thin adoption of fully automated pipelines suggests the next competitive battleground is orchestration, governance and review tooling rather than raw code generation. The report also notes that two-thirds to three-quarters of the sampled companies sell software, so the figures may skew toward technology-forward firms.