By Bruno Müller | 11 November 2026
(Blog prepared with ChatGPT, IHK, Lectures at Hochschule München
Dear Charles,
for our next discussions:
Let us talk today about AI Agents, processing tasks in Agent Swarms.
AI agents can help businesses coordinate work and automate tasks. Their value depends on reliable information, clear decision limits and effective human oversight.
As Ingenieur and entrepreneur, I see considerable opportunities in agentic AI. I also see a systems challenge: when several agents work together (co-create), one agentic mistake can become another agentic starting point.
My central takeaway is simple: Autonomy needs clear guardrails.
AI agents and Multi-Agent Systems?
An AI Agent is software that uses an artificial intelligence model, often a large language model (LLM), to select actions and use tools towards a goal. Depending on its permissions, it might retrieve information, draft correspondence or update’s a business system.
In an AI-Multi-Agent system, several agents contribute to a shared task. One possible design uses an Orchestrator to delegate work and combine results.
The degree of independence varies: some systems follow predefined workflows; others allow agents to choose their next steps. [1]
That distinction matters. A system that prepares a draft needs different controls from one authorised to make a binding commitment.
EXAMPLE No1:
AI Agents in business – The Sales-Office.
Imagine a sales office using three specialised agents:
– An email agent sorts enquiries and identifies customer needs.
– The correspondence agent prepares replies and quotations.
– Telephone agent handles routine enquiries and records requests.
– An AI orchestrator allocates tasks, coordinates handovers and brings the results together.
This could reduce repetitive administration and free people to focus on customer relationships. But consider what happens if the email agent misreads a request. The next agent might prepare a quotation using the wrong assumptions, and the telephone agent might repeat the error.
A polished response can still rest on incorrect information.
For me, this is the central challenge of AI workflow automation: assess the reliability of the complete process, including the handovers between agents.
Why prompt engineering is only part of the answer:
Good prompts help define tasks. Reliable operation also requires clear responsibilities, dependable data, appropriate access permissions and checks on results.
In the sales-office example, I would distinguish between preparing a quotation and approving an exceptional discount. An agent might draft both, while a designated person approves the commercial commitment.
This is human-in-the-loop AI in practical terms: a person has the information, authority and time needed to make a meaningful decision.
The orchestrator needs scrutiny too. Its coordinating role does not make its decisions automatically correct. Logs, review points and independent checks should make mistakes easier to detect and contain.
AI governance: five questions before deployment
For founders and business leaders, I would turn responsible AI into five operational questions:
- Who may decide what? Define the actions each agent can take and the commitments requiring human approval.
- Which information can be trusted? Identify approved sources and how conflicting or outdated data will be handled.
- Who checks the results? Assign responsibility for quality, exceptions and ongoing monitoring.
- When must the system stop? Set escalation rules, spending limits and stop conditions that the system can enforce.
- Who is accountable? Name the people responsible for the process and for responding when something goes wrong.
My preferred starting point is one bounded business process with measurable value. Track time saved, error rates, total operating costs and the amount of human rework. Expand autonomy only when the evidence supports it.
EXAMPLE No2:
Military AI: higher stakes, stronger human control
The coordination principle can also be imagined across aircraft, drones and robotic systems. In military applications, however, a mistaken decision can directly endanger human life.
My view is that effective human control, clear accountability and enforceable operational limits must be central design requirements. Human oversight must give people a real ability to understand, intervene and stop an action.
Developers, organisations, policymakers and affected communities have different responsibilities in deciding where autonomy is acceptable.
„LibreChat“ and open-source AI
[LibreChat](https://www.librechat.ai/) is an open-source AI platform that can be self-hosted. Its agent features let users combine configured models with instructions, files and tools. [2][3]
I was told the comparison with a Linux of AI an interesting metaphor for openness and freedom to shape a system. Whether LibreChat will achieve comparable significance remains an open question.
Self-hosting also requires careful choices. If a deployment connects to external model providers, data may still leave the hosting environment. The privacy implications depend on the complete configuration and the services used.
Frequently asked questions
What is the difference between AI agents and agentic AI?
An AI agent is an individual software component that can select actions towards a goal. Agentic AI describes the broader approach of giving AI systems some independence in carrying out tasks.
What is AI orchestration?
AI orchestration coordinates tasks, tools and information across a workflow. An orchestrator may delegate work to specialised agents and combine their results.
Can AI agents work without human approval?
They can be configured to perform permitted actions automatically. My recommendation is to match that freedom to the consequences of error, with explicit approval requirements for consequential commitments.
Where could AI agents create value in your business?
My starting question would be:
Which recurring process could benefit from AI agents, what improvement would justify the effort, and where must human judgement remain decisive?
Those answers provide a practical foundation for business automation that people can trust.
Sources
[1] [Anthropic: Building effective agents](https://www.anthropic.com/research/building-effective-agents) — architectural principles; published December 2024.
[2] [LibreChat: Agents documentation](https://www.librechat.ai/docs/features/agents).
[3] [LibreChat: Local installation](https://www.librechat.ai/docs/local).
*Sources checked on 11 November 2026.*