Hedwig: AI coding agent that learns when to work autonomously

Coding agents are getting smarter, but developers still struggle to decide how much autonomy to give them. A formative survey of 21 software engineers, conducted by researchers at the University of Washington, revealed that developers are frustrated with existing approaches like static permissions or instruction files. These methods can't adapt to shifting preferences across tasks and over time – an agent that's trustworthy for refactoring might wreak havoc when left alone on a new codebase. The survey confirmed that developers want their AI assistants to earn trust gradually rather than either being completely free or constantly supervised.

Enter Hedwig, a CLI-based coding agent that dynamically adjusts its own autonomy level by learning from developer-agent interactions across sessions. Instead of a fixed global configuration, Hedwig builds an evolving set of behavioral guidelines from the developer's decisions and feedback. When the agent has consistently demonstrated reliability on certain tasks, it relaxes oversight; when it ventures into unfamiliar territory, it tightens control and asks for more input. This approach – accepted at ACM CAIS 2026 – promises to reduce friction and cognitive load for developers, letting them focus on higher-level work while still maintaining local oversight. Hedwig represents a shift from static permission models to dynamic, trust-based autonomy in AI-assisted coding.