Architecture | Enterprise execution protocol

Isolated blank-slate scripts fail in production. Enterprise RAG, long-context preload, and cowork workspaces fail for the same reason: they are libraries and files, not live ERP state. EGI deploys one Digital Nervous System—organizational memory as an index over the ERP, not a corpus. The common context layer is the context graph: continuously refreshed enterprise state. The dual-process loop is the execution graph: intent through completion. Agents hydrate at execution time.

Layer 01

Common Context Layer

· global workspace · lean index · context graph

anti-pattern ·: Not enterprise RAG. Not an embed-everything corpus or session-wide vector preload.
design rule ·: Minimal multi-tenant organizational index—lean, tenant-scoped, synchronized pointers and normalized state for JIT bind at inference and tool-call checkpoints only.
context ·: Cognitive model: shared mental blackboard—specialized regions coordinate through one workspace, not a shared document warehouse.
execution ·: Sits under every agent, above siloed ERP, WMS, TMS, MES, treasury, banking, and data warehouse feeds. Small normalized surface area: entity handles, policy refs, workflow position, legal entity—enough to route what to fetch and where to post, not a document store.
output ·: Purchasing posts a PO → index writes → Warehouse adapts fulfillment. Agents hydrate at Layer 03, not at session start. No agent-to-agent JSON chat.

Layer 02

Individual Agents

· specialized functional cortices · execution graph

context ·: Not generalist chatbots—each agent is a dedicated module with an internal dual-process loop.
system 1 ·: System 1 · neural inference — parses messy inputs (email threads, supplier PDFs) for semantic intent.
system 2 ·: System 2 · symbolic codegen — writes deterministic, rule-bound code against enterprise systems.

runtime · per-agent dual-process loop

INFERENCE

System 1 · understand

CODEGEN

System 2 · execute

WORKFLOWS

Complete

Layer 03

Just-In-Time Retrieval

· hippocampal gate · JIT bind

context ·: Active memory is bounded; the hippocampus lazy-loads targeted fragments on task trigger—not at login.
retrieval model ·: Inference intent + tool schema → bind target on the Layer 01 index → one record or rule slice into the window.
execution ·: Lightweight agent footprint. At policy checkpoints, the Hippocampal Gate queries the index and binds only what that step needs.
output ·: Pulls the minimum slice (e.g. Vendor X bulk discount clause), runs inference or tool call, flushes working memory for the next step.

Protocol surface · versioned state bind · typed transitions · policy gate · authoritative read-back · resumable recovery