arXiv cs.AITuesday · May 26, 2026FREE

Context: Proactive Goal-Directed Intelligence via Composable Sandboxed Programs, Declarative Wiring, and Structured Interaction

Context, presented in an arXiv paper, is an intelligence layer within the Magarshak Architecture designed to shift from reactive query-response chatbots to proactive, goal-directed agents. This system aims to advance shared tasks autonomously, reducing the need for continuous user prompts. Its design integrates three core mechanisms. First, write-time context assembly utilizes Groker agents to precompute enriched, typed attributes, forming interaction context as a deterministic function of graph state. This approach ensures context blocks are byte-identical across turns between semantic changes, facilitating near-100% KV-cache reuse and potentially reducing per-turn LM costs. Second, the architecture incorporates composable sandboxed wisdom programs. These are LM-generated imperative programs, managed as a governed library, which are declaratively wired to specific goal types via typed stream relations. They are composed through phase ordering and executed at interaction time without requiring additional large language model calls. Third, proactive goal stream state machines are employed to drive conversations toward terminal states. These machines inspect the current graph state and emit structured interaction content, such as option arrays, governance affordances, or clarification prompts, without awaiting explicit user input. The paper also formally proves a "Context Stability Theorem," which bounds the per-turn LM cost based on semantic changes, highlighting the architecture's efficiency benefits.