level: technical
context is the intelligence layer of the magarshak architecture. it moves beyond reactive chatbots by using proactive agents that work on shared tasks without waiting for user input. the system has three main parts that work together. first, write-time context assembly uses groker agents to precompute typed attributes. this makes interaction context a pure function of graph state. context blocks stay byte-identical between turns until a semantic change happens, which allows near-100% reuse of key-value caches.
second, composable sandboxed wisdom programs form a governed library of imperative programs made by language models. these programs are wired to goal types through typed stream relations and composed using phase ordering. they run at interaction time without needing more language model calls. third, proactive goal stream state machines drive conversations toward terminal states. they check graph state and send structured interaction signals to keep tasks moving forward.
the design focuses on efficiency and goal completion. by precomputing context and reusing caches, the system reduces redundant computation. sandboxed programs let agents act without constant language model queries, which saves resources. the state machines ensure agents stay on track toward defined goals. together, these mechanisms create a system that can handle complex, multi-step tasks with less latency and more reliability than traditional chatbots.
why it matters: this approach can make ai assistants more autonomous and efficient, reducing the need for constant user input and lowering computational costs in real-world applications.