Kilo Code says developers adopt coding agents through a ladder of trust rather than simply switching to whichever model leads a benchmark. The company based its analysis on traffic through a gateway that has processed more than 40 trillion tokens across over 500 models.
The gateway serves Kilo’s command-line, VS Code, JetBrains and cloud-agent tools. The presented dataset excludes anonymous and internal users, but it still represents customers of one coding platform rather than the entire developer population. Kilo also allows users to bring their own inference, so the gateway does not capture every session.
The proposed ladder starts with low-risk assistance such as autocomplete and chat, then moves toward code-changing agents and coordinated autonomous work. If an early system is slow, selects the wrong file or makes an unexpected edit, users may never grant it authority for higher-level tasks even when its underlying model scores well on public tests.
Kilo identified three recurring failure points: constructing the right context, routing work to an appropriate model and collecting useful feedback. Signals become less immediate as autonomy increases, because a long task may not reveal a bad decision until much later. The practical implication is that teams need instrumentation around actions and outcomes, not only aggregate token use or benchmark scores.