[REVISION HISTORY]

AI adoption and the evolution of software engineering

Updated 34 times since CLSTR started tracking revisions of this situation.

What changed

2026-10-25 06:44 UTC → 2026-10-26 14:02 UTC · added ~~removed~~

Since April 2026, AI agents have transitioned from design pilots to core functions in analytics, energy, finance, and education. While usage is widespread, application remains focused on writing code. At Google, engineer-approved AI-generated code rose to 75% by April 2026. Recent data shows a deepening dependency: a JetBrains survey of 15,000 developers found AI agents fully generate approximately 47% of code, with 21% of developers writing 0% of their code without AI assistance. Emergent founder Mukund Jha reported that approximately 90% of his company’s shipped code now has ‘no direct human touch.’ However, rapid adoption has introduced significant security and operational risks. The rise of ‘vibe coding’ has led to unmanaged dependencies and expanded attack surfaces. ~~Economic efficiency is also being questioned; research suggests only 18 cents~~ Research from Bloom Security indicates that AI agents frequently install packages with limited oversight, with developer machines seeing upwards of ~~every dollar invested~~ 350 agent-installed packages per month. A specific vulnerability in ~~AI-supported development reaches production, with the remainder spent on bug fixes and rewrites.~~ npm Trusted Publishing could allow attackers to publish malicious packages by reclaiming GitHub repository slugs. Furthermore, ~~43% of AI-generated code requires manual debugging, as it can require up~~ AI coding tools have failed to ~~four times more maintenance than manually written code.~~ meet basic web accessibility standards; a study by AudioEye found that 15 generated sites from major providers failed even the most fundamental Level A compliance. A ‘verification crisis’ has emerged as ~~the~~ a primary bottleneck. While ~~AI-powered tools have increased completed tasks by 21%, they have driven~~ AI can generate code at massive scales—with some research suggesting a ~~91%~~ 741% increase in code review time. Developers are shifting from active design to supervisory roles, struggling to comprehend or debug AI outputs. Recent data suggests a discrepancy between increased activity and actual delivery; while AI agents have production—release velocity has increased ~~lines of code~~ by ~~30% and pull requests by 23%, there has been no statistically significant rise in resolved Jira issues or epics.~~ less than 20%. This ~~gap~~ is exacerbated by a lack of independent verification; when AI agents generate both implementation and tests, the tests often merely reflect the same underlying logic errors. This creates a ‘context tax,’ where the time and tokens required to locate files, trace dependencies, and prevent regressions significantly delay implementation. Additionally, a trend termed the ‘hollow ~~orchestrator,’ where~~ orchestrator’ suggests developers may become faster at approving work without improving their ability to judge quality, potentially leading to skill atrophy.

Versions

Only revisions since CLSTR began indexing content versions appear here. Select a version to see what changed compared to the one before it.