level: technical
current ways to describe ai agent architectures look at only one dimension. industry guides from anthropic, google, and langchain focus on execution topology, or how data moves. cognitive science surveys focus on what the agent does. neither view alone can tell apart systems that share a topology but have different behaviors. for example, the same orchestrator-workers setup can run plan-and-execute, hierarchical delegation, or adversarial verification. these patterns have different failure modes and trade-offs.
the proposed framework adds a second axis. the cognitive function axis has seven categories: context engineering, memory, reasoning, action, reflection, collaboration, and governance. the execution topology axis has six structural types: chain, route, parallel, orchestrate, loop, and hierarchy. together they form a 7 by 6 grid. this grid maps 27 named patterns, with 13 receiving original names. the two dimensions make it easier to compare designs and spot gaps.
the classification helps builders choose the right pattern for a task. it shows that patterns with the same topology can serve very different cognitive goals. the framework also reveals empty cells where no known pattern exists, pointing to possible new designs. by separating what an agent does from how it is wired, the matrix gives a clearer map of the design space. this can guide both research and practical system building.
why it matters: it gives ai developers a structured way to compare agent designs and avoid mismatches between architecture and cognitive needs.