Multi-Agent-Koordination
Multiple specialists, one shared goal — coordinated by the Orchestrator, secured through permission levels, connected via secure communication channels.
Konzept
An employee responsible for accounting, HR, and engineering simultaneously cannot do any of them well. This applies to people — and equally to AI agents. Separate specialists are not a technical luxury, but a regulatory and architectural necessity.
The sales agent cannot see HR data, the HR agent cannot see engineering drawings. Each agent has access exclusively to the data of its own domain — architecturally enforced, not by policy.
DSGVO Art. 25 • ISO 27001 A.9 • BSI IT-Grundschutz ORP.4
Whoever books must not also approve. An agent that processes customer data, financial accounting, and personnel files violates the principle of separation of duties. Separate agents technically implement organisational separation lines.
GoBD Tz. 107 • HGB §239 • DSGVO Art. 5
With a 14,000-token context window, a generalist uses most of it for its instructions. Specialists use only 17–22% for their prompt — the rest is available for your query, memories, and tools.
→ Architektur: 7 PrinzipienEach agent produces its own, separate log. During an audit you can immediately see: welcher Agent, wann, welche Daten, welches Ergebnis. No disentangling a monolithic log.
GoBD Tz. 37 • ISO 9001:2015 Kap. 7.5 • DSGVO Art. 30
If an agent malfunctions, it affects only its domain. The other agents continue working undisturbed — like fire compartments in a building.
ISO 27001 A.16 • BSI 200-4 Notfallmanagement
The engineer talks to the engineering agent, accounting talks to the finance agent. Each user sees only the data relevant to their role — Privacy by Design at the architecture level.
DSGVO Art. 25 (Privacy by Design)
Fazit: Multiple specialized agents are the consistent implementation of data protection, separation of duties, and quality assurance — just as you would not assign a single person to accounting, HR, and engineering simultaneously in your organisation.
The Orchestrator coordinates who uses the GPU and when. Agents communicate through secure channels and share files — without access to each other's workspace.
Coordination
Inter-agent messaging, shared files, and architecturally enforced data privacy.
Agents of the same connector family share a communication channel. Messages to the family are automatically routed to the responsible agent.
Maximum 3 back-and-forth rounds between two agents per request. Prevents infinite loops and uncontrolled token consumption.
Each agent has its own workspace — like a locked desk. Only deliberately shared files are visible to others.
Reports, exports, BOMs — everything other agents need. The agent decides what to share.
Notes, calendar, contacts, todo lists, profile databases. No other agent has access — architecturally enforced, not by policy.
Beispiel: The engineering agent reads a DXF file from the engineer's workstation, creates a BOM, and places it in public/. The finance agent reads the BOM and calculates material costs — without access to engineering notes.
Security
Three trust levels determine which tools an agent may use. Like an operating system kernel.
Lowest risk. The agent may only read: retrieve memory, view configuration. No access to external systems, no write access.
Read access to external systems. SQL SELECT, REST GET, read files. No write access to production data. Default ring for most agents.
Full access including write operations. Requires explicit authorisation by the administrator. Human-in-the-loop for critical actions.
Practical Examples
Each agent is individually configured for its task — with its own skills, connectors, and knowledge base.
Berechnet Tragwerke, optimiert Stahlprofile, prüft Normen. Creates BOMs as draft — the engineer reviews and approves.
Accesses ETA, DATEV, or other accounting systems. Delivers balances and cash flow in seconds. Prepares bookings — the accountant reviews and approves.
Monitors inventory via ERP interface, detects shortfalls. Creates requisition — Einkäufer gibt Bestellung frei.
Reads machine status, optimises order sequence, creates shift handover reports. Suggests sequence — production manager approves.
Checks vacation requests against team calendar, tracks onboarding, warns about contract deadlines. Recommends decision — Führungskraft genehmigt.
Scans contracts for deadlines, generates NDA drafts from templates, reports regulatory changes. Creates draft — the lawyer reviews and approves.
Receives monitoring alerts, enriches tickets, monitors license expirations. Priorisiert Incident — the admin reviews and escalates.
Individually configured for your department, your systems, your workflows.
Jetzt besprechenThese are examples. AIMOS agents are individually configured per company — every department, every interface, every workflow. More about the agent architecture.