Context: Proactive Goal-Directed Intelligence via Composable Sandboxed Programs

Forum topic · 小凯 · 2026-05-27

Summary

This post summarizes the arXiv paper "Context: Proactive Goal-Directed Intelligence via Composable Sandboxed Programs, Declarative Wiring, and Structured Interaction" by Gregory Magarshak. Context is the intelligence layer of the Magarshak Architecture, replacing reactive query-response chatbots with proactive goal-directed agents that advance shared tasks without waiting for user prompts. The design combines three mechanisms: write-time context assembly, where Groker agents precompute typed attributes so context is a deterministic pure function of graph state, enabling near-100% KV-cache reuse across turns; composable sandboxed wisdom programs, LM-generated imperative programs wired declaratively to goal types via typed stream relations and executed without further LM calls; and proactive goal stream state machines that emit structured interaction content such as option arrays and clarification prompts. The paper proves six formal results, including a Context Stability Theorem bounding per-turn LM cost, a Proactive Dominance Theorem, and a Cross-Platform Vote Consistency Theorem. It is implemented in the open-source Qbix / Safebox / Safebots stack.

Paper Overview

Field: NLP Author: Gregory Magarshak arXiv: 2505.21641

Abstract

We present Context, the intelligence layer of the Magarshak Architecture, which replaces reactive query-response chatbots with proactive goal-directed agents that advance shared tasks without waiting for user prompts. The architecture rests on three mutually reinforcing mechanisms. Write-time context assembly precomputes enriched typed attributes via Groker agents, assembling interaction context as a deterministic pure function of graph state; context blocks are byte-identical across turns between semantic changes, enabling near-100% KV-cache reuse. Composable sandboxed wisdom programs form a governed library of LM-generated imperative programs declaratively wired to goal types via typed stream relations, composed via phase ordering, and executed at interaction time without further LM calls. Proactive goal stream state machines drive conversations toward terminal states by inspecting graph state and emitting structured interaction content (option arrays, governance affordances, clarification prompts) without awaiting user input.

Formal Results

The paper proves six formal results:

1. Context Stability Theorem — bounds per-turn LM cost as a function of semantic change rate. 2. Program Composition Correctness Theorem. 3. Declarative Wiring Soundness Theorem. 4. Proactive Dominance Theorem — proactive agents weakly dominate reactive agents on expected turns-to-terminal-state. 5. Coordination Overhead Elimination and Quality Preservation — establishes Pareto improvements in multi-participant goal chats. 6. Cross-Platform Vote Consistency Theorem.

Implementation

Context is implemented in the open-source Qbix / Safebox / Safebots stack.