NLP Expert Group Meeting - AI Agents

From Assistant to Agent: When AI Arrives in the Process

The term "AI agent" is used quickly, yet it often remains unclear whether everyone means the same thing. In my webinar I therefore start with a precise definition. A reactive assistant answers requests and delivers content "alongside the process". A process-integrated agent is designed to work within the workflow: it responds to events, pursues goals across multiple steps, uses tools and interacts with systems. What matters is not maximum autonomy but reliable execution within clear guardrails, so that results remain traceable.

What are AI Agents? I guess we know, but do we agree? Reactive Assistants vs. Process Integrated Agents

A Spectrum Rather Than Black and White: Control vs. Freedom

Agents are rarely an either/or. I describe them as a spectrum between predefined workflows and fully autonomous systems. In practice the productive range often lies in between: depending on risk and complexity, "more freedom" is not automatically better. An agent is useful when it has enough room to get work done but enough control to remain stable, technically and organisationally.

The Spectrum of Agents — Control vs. Freedom

Modularity as a Principle: Building Blocks Rather Than a Monolith

Another focus is modularity. When agents are built as monolithic all-rounders, they quickly become difficult to test and difficult to steer. I therefore think in "building blocks": clearly bounded components that can be combined without every extension destabilising the overall system. This is both an architecture principle and a team principle: capabilities remain interchangeable, responsibilities stay clear and further development becomes more plannable.

Building Blocks — The Power of Modular Agents

Role Model TACO: How Agents Act in Collaboration

To make functions in agentic systems tangible I use a role model (TACO). It helps distinguish agents by their function in the value chain: from focused "Askers" for single goals, through "Automators" for recurring tasks, to "Collaborators" for human-agent teamwork and "Orchestrators" that coordinate complex end-to-end processes across multiple systems. As orchestration grows, integration and operational requirements typically rise as well, making it increasingly necessary to design communication and control carefully.

Focus on single goals, broken into simple steps. Are the easiest to manage. Are for human-agent teamwork, acting like teammates, not just tools Manage complex, end-to-end processes across systems

Multiple Agents, One Core Problem: Communication

As soon as multiple agents work together, communication becomes the central system question. I draw on the "Agent Communication Trilemma": efficiency, portability and versatility are in tension with one another. An approach that is very efficient for a specific environment often loses portability. A highly portable approach is frequently less optimised. And maximum versatility can increase the cost of standardisation and operations. The practical consequence follows: agents need an explicit language for tasks, states and results, appropriate to the system landscape and the operating model.

The Agent Communication Trilemma EFFICIENCY — PORTABILITY — VERSATILITY

Protocols as Bridges: MCP, Agent Protocol and Chat-Like Approaches

I map several protocol directions that address precisely this gap between the model and the "outside world". The Model Context Protocol (MCP) stands out as an approach to standardising tool and context integration. Agentprotocol.ai describes work more strongly via tasks, steps and artefacts, in clear, process-oriented units. And chat-inspired protocols like the LangChain Agent Protocol lean more toward conversation and runtime concepts that many agent systems already use.

Model Context Protocol (MCP) — Anthropic's approach Just Tasks, Steps and Artifacts? Langchain Agent Protocol — A chat-inspired approach

AGORA: Protocols on Demand, Natural Language Only When Necessary

As a conceptual addition I introduce AGORA: an approach that allows different protocol forms and uses appropriate structures depending on the situation, up to and including "Protocol Documents" that can be referenced. The guiding idea is pragmatic: natural language is flexible but hard to validate and version. It is therefore treated as a last resort rather than the default transport format for precise states.

Natural language as a last resort Agents negotiate which protocol to use

A Target Architecture of Our Own: Closed, Event-Based, Observable

In closing I condense the content into a practice-oriented target picture: a closed, event-based system that enables proactivity while consistently keeping humans in the loop. Transparency is a basic requirement: actions and events must be observable so that debugging, auditability and operational security become possible. Workflows are predominantly predefined, with AI acting mainly on routing and flexible execution within the guardrails; thread and run concepts serve as a structural orientation for the runtime.

Transparency — Every action and event can be observed Predominantly predefined workflows with AI routing Event-based system allowing for proactivity — Human in the Loop