The Developer as Curator: A Canvas for Coding Agent Context

Abstract

The quality of AI-generated code depends on the context that coding agents automatically receive as input. Today's coding agents do not expose that context to the developer. When retrieval fails, the developer cannot distinguish whether a wrong output stems from flawed reasoning or from incorrect or missing input. We describe how coding agents retrieve context internally and derive three failure modes: wrong context, missing context and stale context. Existing mechanisms for managing context (compaction, session reset, and sub-agent delegation) are automated responses to token limits that give developers no way to correct these failures. We present Monet, a canvas-based development tool that positions the developer as curator of the context the agent receives. Monet is guided by two design goals: context observability, letting developers see every piece of gathered context and its current state, and context curation, letting developers add, remove, and refresh context through direct manipulation. On Monet's canvas, each piece of context is a node that developers can remove, add, or refresh to correct the agent's input before re-prompting. Case studies show that each mechanism catches retrieval errors that autonomous agents cannot self-correct. Monet opens a new design space for human-AI collaboration in agent-assisted programming, shifting the developer's role from prompt author to context curator.