An increasingly common approach to AI systems is to divide work between specialised agents.
One agent interprets requirements, another plans work, another performs implementation activities and another validates the results.
This decomposition has advantages. Individual capabilities can be developed, tested, maintained and replaced without requiring a single model to perform every function.
But it introduces a significant problem: preserving context between agents.
When information passes through a sequence of models, important details can be lost. Assumptions become detached from their sources, decisions are repeated and subsequent agents may interpret the same requirements differently.
Passing the entire conversation history to every agent is one possible response, but it becomes increasingly expensive and difficult to maintain as the amount of information grows.
A more useful approach may be to separate the information the agents work with from the agents themselves.
A shared context system could maintain requirements, entities, relationships, decisions, policies, evidence and the current state of each process. It would also record which information is authoritative, which has been inferred and which remains uncertain.
Each agent would retrieve the context relevant to its responsibility, perform its work and return structured results to the shared system.
For example, a requirements agent might identify an ambiguity in an approval process. That ambiguity would be recorded against the relevant requirement, with its source and the decision needed to resolve it.
A configuration agent could recognise that the decision remains outstanding. A testing agent could subsequently verify the approved behaviour without needing to reconstruct the original discussion.
The context would need to be maintained through explicit identifiers, versioned decisions, access controls, event histories and references to original evidence.
Short summaries may still be useful for efficient retrieval, but they should not replace authoritative records. Agents must also be able to recognise when earlier context has become outdated or contradictory.
Existing protocols such as MCP and Agent2Agent provide foundations for accessing tools and coordinating agents. They do not, by themselves, guarantee a persistent shared understanding of the business information being processed.
That semantic layer is a separate architectural responsibility.
The potential benefit extends beyond efficiency. Agents could be updated, replaced or assigned different responsibilities without losing the history and decisions on which their work depends.
The long-term maintainability of multi-agent systems may depend as much on how context is preserved as on the capabilities of the individual models.