Architecture field guide · 01

Agents execute.People decide.

A practical operating model for organizations where specialized AI agents work privately at machine speed, surface what matters, and learn from every human decision.

Explore the architecture →Run a scenario
Brain, memory, cache and dreaming connected in a learning loop Collaboration planeExecution planeKnowledge plane
PRIVATE → RELEVANT → SHARED → DECISION → PRIVATE → LEARN
Private by default→Visible when meaningful→Human authority where needed→Learning over time

Watch the architecture

See the operating model in motion.

A guided walkthrough of private agent execution, visible human–agent collaboration, the communication router, and the memory, caching, and “dreaming” mechanisms that make the organization learn.

01 Two communication layers02 Policy-driven routing03 Memory, cache & dreaming04 Enterprise use cases
A human and AI agents collaborating above a connected specialist network, surrounded by memory, caching, dreaming, and governance symbols
Human judgment above · Machine-scale coordination belowKnowledge connects both

The three-plane model

Three planes.
One learning system.

The architecture separates execution, collaboration, and knowledge because each has a different audience, cadence, and tolerance for noise. Governance crosses all three.

01 / VISIBLE

Collaboration plane

Humans and agents share findings, disagreements, approvals, and decisions in familiar workspaces.

SlackTeamsDiscordBuzz
02 / PRIVATE

Execution plane

Specialized agents delegate, calculate, validate, and use enterprise tools without flooding human channels.

A2AMCPAPIs
03 / OVER TIME

Knowledge plane

Memory preserves, cache reuses, and governed “dreaming” reconciles what the organization has learned.

MemoryCacheDreaming

Reference model

The hybrid operating model, in one view.

Trace the collaboration, execution, and knowledge planes through a bidirectional communication router—from private machine work to shared decisions and back into learning.

The hybrid human–agent operating model showing the collaboration, execution, and knowledge planes, communication router, promotion triggers, and operational loop
Hybrid human–agent operating modelOpen diagram at full size ↗

The routing loop

Meaning moves up.
Authority moves down.

Private
execution
→
Detect
relevance
→
Shared
context
→
Human
decision
→
Execute
& learn
Try the routing simulator →

Six building blocks

From conversation
to controlled action.

Each block has one job. Together they let organizations balance speed with visibility, autonomy with authority, and learning with control.

01

Collaboration layer

The workroom where people and agents develop shared understanding.

Open layer →
02

Communication router

The policy engine that decides what stays private and what becomes shared.

Test the router →
03

Agent network

Specialists that discover and delegate to each other through A2A.

See the network →
04

Systems & tools

ERP, CRM, data, code, search, and workflows accessed through MCP and APIs.

See interfaces →
05

Governance

Identity, permissions, policy, audit, observability, and cost controls.

Explore controls →
06

Knowledge & learning

Memory, caching, and dreaming make tomorrow’s execution better than today’s.

Follow the loop →
Human working visibly with a group of AI agents

Design intent

Expose work products, not hidden reasoning.

Visible agents should share findings, evidence, confidence, risks, questions, and recommendations. Hundreds of internal tool calls and intermediate messages stay below the surface.

Compare the two layers →

The organizing principle

“Agents communicate programmatically by default and socially when information becomes organizationally relevant.”