Orchestrator
2026-09-06 · 6 min read
Key takeaways
- Agents are typed sequences of steps, not chat prompts.
- A Orchestrator coordinates agents across several systems.
- Human approval is a step, drawn and colored like the others.
ERP, PLM, CRM and documents keep provenance and version.
Typed steps
Each agent step declares its kind: an API call to a system service, an AI reasoning step with its model, or a human approval.
That is what makes an agent reviewable: you can read what it will touch before you let it run.
The worker as an alter ego
A Orchestrator maps to a business role — buyer, planner, pricing manager — and holds the agents, workflows and authority of that role.
It can propose an activity list for its agents and for people, but it never approves its own proposal.
Technology, not systems
An agent often spans several systems, so it is classified by the technology it was built on: Claude, OpenAI, Gemini, n8n, SAP BTP or native.
Provenance stays visible, especially for agents imported from market LLMs.
FAQ
Frequently asked questions
Why does “Typed steps” matter?
Each agent step declares its kind: an API call to a system service, an AI reasoning step with its model, or a human approval. That is what makes an agent reviewable: you can read what it will touch before you let it run.
How does “The worker as an alter ego” work?
A Orchestrator maps to a business role — buyer, planner, pricing manager — and holds the agents, workflows and authority of that role. It can propose an activity list for its agents and for people, but it never approves its own proposal.
What should a team know about “Technology, not systems”?
An agent often spans several systems, so it is classified by the technology it was built on: Claude, OpenAI, Gemini, n8n, SAP BTP or native. Provenance stays visible, especially for agents imported from market LLMs.