How should agent capabilities be organized into functional layers?

In the same talk, Dubberly lays out software-agent capability as "a stack of layers." At the base, "LLM-based chatbots are a simple form of agent" — citing ChatGPT, Claude, and Cursor in the US and ChatGLM and Ling-1T in China — that "take an input and transform* it to return an output." Agents "often have a second input — a system prompt, describing the agent's role, tone, and constraints," and "may draw on information sources outside the LLM through a process called Retrieval Augmented Generation (RAG)." Scaling up, he argues "every physical process and virtual process (that matters) will be overseen by a monitoring agent (a.k.a. Unix daemons), alerting supervisor agents when errors occur or are predicted," and that "the supervisor agents will ladder up in a management tree to an orchestrator agent at a human touch-point (though humans may drill all the way down if they like)."

The stack has two distinct axes: a capability axis (chatbot to system prompt to RAG, each adding an input that narrows or extends what the base LLM transform can act on) and a scale axis (single monitoring agents aggregating into supervisor agents aggregating into one orchestrator agent at the human interface). Dubberly pairs the scale axis with a warning rather than an endorsement, quoting Xerox PARC researcher Rich Gold's 1995 question about "intelligent houses": "How smart does your bed have to be, before you are afraid to go to sleep at night?" — flagging that a management tree of agents "everywhere" is a design problem as much as a capability one.

The scale axis parallels but does not duplicate Does machine agency exist on a spectrum rather than binary?: Rammert's spectrum rates how much initiative a single artifact takes, from passive to cooperative, while Dubberly's stack describes how many agents compose into a hierarchy regardless of any one agent's initiative level. It also gives mechanism to Can AI agents truly operate above the loop at enterprise scale?, which places humans above or within AI-first workflows without specifying what sits between the workflow and the human; Dubberly's orchestrator-agent-at-the-touch-point fills that gap. Against How does agency split between humans and AI systems?, the stack adds that supervision itself can be agentic — a supervisor or orchestrator agent, not a human, may hold the oversight role the archetypes assume belongs to a person. The companion note, Is design for agents becoming the next major design paradigm?, is the paradigm-level claim this stack is offered in service of.

The excerpt gives no mechanism for how an alert actually propagates from a monitoring agent to a supervisor agent, no accuracy or false-positive data for prediction ("errors occur or are predicted"), and no account of what happens when an orchestrator agent itself errs. "Billions of them" is a rhetorical projection, not a sourced forecast. The stack should be read as a proposed design vocabulary for talking about agent hierarchies, not a validated architecture or a measured prediction of how many agents will exist.