Architecture Overview

EAK divides agent infrastructure into three cooperating layers: identity, memory, and action. Together they let agents execute tasks safely, continuously, and controllably in real production environments.

A reviewable infrastructure loop

1. Intent: A user, business system, or another agent starts a task.
2. Identity: GenAuth confirms the human user, Agent Profile, authorization scope, and MCP access boundary.
3. Memory: GUMem retrieves user profile, conversation history, action traces, and search context.
4. Action: Web Agent or MCP tools operate on real webpages and third-party APIs.
5. Review: The system preserves authorization, context provenance, and action traces for later inspection.

These five steps are not just a diagram. They are the default thread that the docs site uses to explain how the stack works.

Layer responsibilities

GenAuth

GenAuth is the identity relay layer. It transforms human ownership into temporary authorization that agents can safely carry. It is not just login. It is the permission boundary for agent work.

GUMem

GUMem is the shared memory layer. It turns conversations, searches, actions, and profile signals into recallable, explainable, and erasable user context.

Web Agent

Web Agent is the action layer. It lets agents connect to the open web through search, extraction, cleaning, browser actions, tracking, and sandboxed execution.

Design constraint

EAK's architecture is not a collection of tools placed side by side. It is an explainable flow that should always answer:

Who authorized the agent, what context shaped the decision, and which action happened?

That question is the production constraint shared by every module in the stack.