Kai Schiller publishes Hugging Face community guide to limiting AI agents

Why it matters

AI-agent products are defined as much by permissions and stopping rules as by model quality. Schiller's guide gives builders a practical test for separating executable workflows from renamed chatbots, while smolagents shows where those limits can be enforced in an implementation.

Kai Schiller published a German-language community guide to AI agents on September 9th, 2026, that strips the category down to its working parts: a model, a goal, a set of tools, memory and controls over what the software may do.

Schiller publishes under the Agenten profile, where he lists interests including AI agents, Model Context Protocol, large language models, automation and open-source AI. His article takes a builder's approach to establishing where a chatbot ends and an agent begins.

Hugging Face labels the piece a community article, so it should not be read as a corporate technical paper or product announcement. Schiller defines an agent through its ability to pursue a goal using a sequence of decisions and external actions.

That distinction matters because "agent" has become a roomy label for products ranging from chat interfaces with a search button to software that can write code, query databases, send messages or change records. Schiller proposes a practical test: does the system break a task into steps, select tools, evaluate the results and decide what to do next?

The loop is the product

Schiller describes a six-stage cycle. The agent interprets a goal, chooses a next step, selects a tool, executes the action, evaluates the output and decides whether another step is required. The run ends when the goal has been reached or the system can no longer identify a useful action.

A train-planning example makes the difference concrete. A chatbot asked for inexpensive travel from Bielefeld to Berlin might explain how to find a ticket. An equipped agent could query a travel service, compare routes by price and duration and produce a recommendation from the returned data.

The language model can remain largely the same in both cases. The architecture around it determines whether the model produces text or directs a workflow.

That framing matches Hugging Face's smolagents documentation, which describes agents as multi-step systems in which a model decision can lead to a tool call and execution.

The framework competes in a busy agent-orchestration market. LangGraph combines agent orchestration with observability, evaluation and deployment tooling. CrewAI emphasizes teams of specialized agents and event-driven workflows, including managed offerings. Schiller's guide supplies no comparative benchmark, cost test or production reliability data for smolagents or those alternatives.

Autonomy becomes a permissions problem

Schiller includes feedback and control among the five components of an agent. The guide describes maximum step counts, cost limits, restricted tool permissions, result checks and explicit approval before consequential actions.

An incorrect chatbot can return an incorrect sentence. An incorrect agent can pass bad parameters to an API, alter a file or send a message. Each extra permission expands the cost of a bad decision.

The guide treats autonomy as a deliberate design choice, ranging from software that prepares suggestions to bounded workflows that use tools and complete multiple steps. A production system still needs a clear stopping condition, a constrained toolset and a route back to a person. Additional model and tool calls also increase operating costs and give mistaken intermediate results more chances to shape later decisions.

MCP connects the tools, with limits

Schiller presents the Model Context Protocol as infrastructure for connecting AI applications to files, databases, internal knowledge and developer tools through a common interface. The specification defines an open protocol built around hosts, clients and servers, with resources, prompts and tools exposed to AI applications.

MCP can reduce the custom integration work required for each service. It supplies connections to tools and data; the application remains responsible for the goal, planning loop, permissions and control policy that govern their use.

The guide also discusses multi-agent systems in which specialized agents divide research, analysis, coding, criticism and coordination. Adding agents increases compute costs, communication overhead and the number of places where errors can enter. For founders, the operational question is whether those extra roles make a workflow easier to evaluate and maintain. The guide provides no evidence that a larger agent group improves performance.

Hugging Face provides the distribution layer

The guide shows how Hugging Face functions as a publishing and distribution venue for technical practitioners outside the company's staff.

Hugging Face was founded in 2016 by Clement Delangue, Julien Chaumond and Thomas Wolf. Hugging Face is based in New York and lists a European headquarters in Paris.

Hugging Face's current Hub documentation says the platform hosts more than 2 million models, more than 500,000 datasets and more than 1 million AI applications, known as Spaces. More than 50,000 organizations use the platform, according to the Hugging Face homepage. Its open-source catalog includes smolagents alongside Transformers, Diffusers, Datasets and other machine-learning libraries.

Hugging Face also operates a paid software and infrastructure business around that catalog. Its current plans list Team and Enterprise access starting at $20 per user per month, while GPU compute starts at $0.60 per hour. Companies use the platform to store private machine-learning assets, run applications and buy access to hosted infrastructure.

Hugging Face announced a $100 million Series C led by Lux Capital, with Sequoia, Coatue, Addition, AIX Ventures, SV Angel and Betaworks among the investors. In August 2023, Hugging Face raised a $235 million Series D at a reported $4.5 billion valuation. Salesforce Ventures led that round, joined by Google, Amazon, Nvidia, Intel, AMD, Qualcomm, IBM and Sound Ventures.

That scale gives even a basic taxonomy practical weight. Builders choosing models, tools and examples on the Hub need a shared vocabulary for what they are assembling. Schiller's answer is deliberately mechanical. The model decides, tools act, memory carries context and controls determine when the process stops or returns to a human.

The guide makes no performance claim and supplies no production reliability data. It offers a design test that founders can apply before calling a workflow an agent: identify the decisions it can make, the systems it can touch, the information it remembers and the point at which a person takes control.