Context Synchronization for Multi-Agent Workflows

Paper — 2026-04-22
ECE 57000 Course Project
Ryan Baker

Paper · Link

Abstract

Large language model agents have demonstrated remarkable capability as autonomous coding assistants, motivating workflows in which multiple agents operate concurrently on a shared codebase. However, existing frameworks treat each agent’s context as private and static: agents reason from symbols, signatures, and dependencies observed at task start, unaware when concurrent edits have invalidated those assumptions. This paper presents CodeCtx, a shared context layer for multi-agent coding workflows. CodeCtx constructs a symbol-level dependency graph over a Python repository, records each agent’s observations as explicit context dependencies, and propagates targeted invalidations when files change.