Context in multi-agent systems and KISMA as a remedy

We have moved from AI dialogs to agentic workflows, connecting specialized agents to handle more complexity and autonomy. But in doing so, we’ve introduced the challenge of how the system components maintain a common understanding – or how it’s called in AI terms, a common context. A paper from the Alibaba group proposes an architectural solution: KIMAs (knowledge integrated multi-agent systems).

The importance of context

Imagine a system where a specialized agent helps a large foundation model to find updated product and price information on the internet. In an early part of the dialog, the user talks about “Dell monitor” and in a refinement request which is handled by the specialized agent, the user only uses the pronoun “it“ to refer to the search intent.

An empirical study by Google "Sufficient Context: A New Lens on Retrieval Augmented Generation Systems" examines the amount of context and its influence on response quality. As you can imagine: less context leads to more hallucination.

Categorizing RAG Responses: Sufficient vs Insufficient Context

Knowledge integrated multi-agent systems

Back to solutions. KISMA proposes as part of the pipeline of a multi-agent system one component that takes over content: the context manager. A router that analyzes requests using a small model with low latency decides, if the content manager is needed or not. If it’s instantiated, it manages the context for all system components tailored to their need.

Simple but very effective in alleviating the lack of content and the hallucinations that result. The paper is well worth reading: KIMAs: A Configurable Knowledge Integrated Multi-Agent System.