<p>Six months into widespread production adoption of AI coding agents, the industry is moving past hype and into pragmatic assessment. A survey of 1,200 engineering teams across Europe and North America, conducted by software consultancy ThoughtWorks, reveals patterns that differ significantly from vendor marketing claims.</p>
<h2>Where Productivity Gains Are Real</h2>
<p>Teams report 35-45% productivity improvements on well-defined, bounded tasks: writing unit tests, generating boilerplate, translating code between languages, and refactoring isolated functions. These are tasks with clear inputs, outputs, and verifiability. The human review burden for these outputs is low.</p>
<p>Bug fixing in well-tested codebases shows a similar pattern — agents are effective when the test suite can validate their output. Where tests are sparse, agent-generated fixes introduce regressions at a rate that erases the time savings.</p>
<h2>Where Agents Struggle</h2>
<p>System design, cross-module refactoring, and novel feature development show much smaller gains and sometimes negative productivity — engineers report spending more time reviewing and correcting agent output than writing code themselves. Complex tasks requiring context from dozens of files and understanding of implicit business constraints remain primarily human work.</p>
<h2>Emerging Best Practices</h2>
<p>High-performing teams have developed a consistent pattern: agents write first drafts, humans review and specify corrections, agents iterate. This works well when humans maintain architectural authority and use agents for execution. Teams that delegate architectural decisions to agents report higher rework rates.</p>
<p>The survey also found that teams with higher test coverage got more value from agents — a finding that's reshaping investment in test infrastructure as a prerequisite for effective AI adoption.</p>