madarX
  • Overview
  • Projects
  • Agents
  • Tasks
  • Workflows
  • Approvals
  • Marketplace
  • Settings
Loading Federated AI…
document · 004 · federationmadarX OS · company agent network
architecture atlasconsole

AI generates. Humans approve. Departments collaborate.

From one AI assistant to a governed company-wide AI workforce.

madarX OS can become the operating layer where every employee and every department has its own team of agents — while cross-department work moves through policies, approvals, queues, memory boundaries, and audit trails.

strategic advantage
vs Hermes
not only personal execution — institutional coordination
vs Paperclip
not only control-plane records — real governed runtime
core moat
federated agent teams with approval-first collaboration
scale path
local-first → team deployment → enterprise network
Current operating modelFederated organization

Departments collaborate through a broker, policy engine, approvals, events, and audit trails — not uncontrolled agent-to-agent calls.

request brokerpolicy engineevent busapproval ledger
One-line product claimpositioning

madarX OS turns individual AI agents into an approval-first, audit-friendly operating system for entire companies.

Department agent networkclick a department
Financedepartment tenant
manager
Finance Controller Agent
memory
budgets · revenue · forecasts
tools
supabase_read · sheet_export · finance_api
approval
Restricted financial data
Budget AnalystRevenue ReporterForecast AgentApproval Controller
Cross-department communication simulatormarketing asks finance, legal asks engineering, support asks ops
approval-required

Marketing → Finance

Risk tier: high. Approver: Finance manager.

  • cross-department boundary
  • confidential sensitivity
Records createdcross_department_requestsactivity_logapproval_requestsaudit_logsdepartment_queue_jobs
Request flowbrokered, not free-for-all
01Request created

Source agent submits intent, business reason, target department, requested fields, sensitivity, and urgency.

02Broker normalizes

Request Broker attaches tenant_id, department_id, actor_id, task_id, correlation_id, and target queue.

03RBAC + policy

Policy engine checks role, department, data owner, scope, risk tier, rate limits, and active approvals.

04Approval decision

Low-risk can auto-approve by policy; sensitive finance/legal/customer/tool actions pause for manager approval.

05Target agent executes

Target department manager assigns a specialist agent with scoped memory and tool permissions.

06Result returned

Only approved fields or artifact links are returned. Raw restricted data can stay inside the owning department.

07Audit + replay

activity_log, audit_logs, approval_requests, tool_egress, and task events preserve the full chain.

Why this is saferboundary design

Agents do not directly scrape another department’s memory or tools. They request work through a broker. The owning department decides what can be returned.

  • Raw finance data can remain inside Finance.
  • Marketing receives approved summaries, not unrestricted tables.
  • Legal opinions require Legal manager approval.
  • External sends and destructive tool calls remain irreversible_external.
Federation architecturesix-layer model
01People + departmentsemployees · department heads · executives · contractors
02Department tenancytenant_id · department_id · workspace_id · memory namespace · budget profile
03Governance planeRBAC · policies · approval rules · rate limits · audit rules
04Coordination planerequest broker · task router · event bus · queues · heartbeat worker
05Intelligence planemanager agents · specialist agents · Hermes · Claude Code · Codex · Ollama
06Data + toolsSQLite/Postgres-ready store · vector memory · files · GitHub · Supabase · email · terminal
Performance simulatorqueues · workers · audit volume
1,440task/hour capacity
69%worker utilization
8mest. queue wait
5,400audit events/hour
Scalability ruleskeep it fast and inspectable
  • Every request carries tenant_id, department_id, actor_id, task_id, correlation_id.
  • Hot permissions, tool manifests and rate counters can be cached.
  • Long work moves to department queues with leases, retries and dead-letter handling.
  • Vector memory retrieval is scoped by tenant, department, role, task and policy.
  • Audit logs are append-only and can be partitioned by company/department/time.
Advantage matrixHermes · Paperclip · madarX OS
DimensionHermes AgentPaperclipmadarX OS advantage
Primary shapePersonal runtime for one operator/session.Company/work control plane and artifacts.Federated AI workforce: employees, departments, agents, approvals, tools, queues, audit.
Department ownershipNot modeled as departments by default.Can model org work, but runtime is external.Each department owns agents, memory namespace, tools, queue, policies and approval manager.
Cross-team workdelegate_task spawns subagents, not company-governed departments.Tasks/governance exist, but not necessarily local execution.Brokered cross-department requests with RBAC, policy, manager approval and audit lineage.
Approval strengthStrong command approvals for personal execution.Governance workflows in control plane.Approval-first at company boundary: cross-department data, tool scopes, external sends, deploys, finance/legal restrictions.
Execution enginesHermes itself runs tools and providers.Adapters phone home; plane can stay separate.Claude Code, Codex, Hermes, Anthropic, Ollama and future MCP tools behind a single governance layer.
Scalability pathGreat single-user/gateway scale.Org data model scale.Department queues, worker pools, event bus, policy cache, memory scopes, audit partitions, multi-tenant cloud path.
Business moatOperator productivity.Structured autonomous company records.Accountable AI labor market inside the company: everyone gets agents, and agents can collaborate safely.
Department war-room demolaunch a campaign across six departments
Finance · approval.required

Finance Controller allows summarized campaign budget but blocks raw revenue exports.

Every handoff creates durable task, activity, approval, and audit context. This is the difference between “agents chatting” and an accountable AI organization.
madarX OS · federated AI teams · document 004company-wide agents · governed collaboration · audit-first