AI coding agents are blowing through budgets — Replit, Kilo Code, and Symbotic explain how they're managing it

Engineering leaders from Replit, Kilo Code, and Symbotic say AI coding agents are now handling much of the day-to-day coding workload, with Kilo Code estimating humans read or write code only about 1% of the time. They described guardrails such as risk-scored pull requests, human review for complex changes, and clearer criteria for security, maintainability, and code quality.

The shift suggests AI coding tools are moving from productivity experiments into core software delivery workflows, but cost control and governance are becoming just as important as capability. Expect teams to build more agent oversight systems, especially for brownfield codebases where product context, legacy constraints, and cleanup work remain hard to automate.