Connected systems change the relationship between an output and the work that follows it. A result can become persistent context, available to other tools and returned to the process that produced it. Over successive iterations, the system begins to work with an increasingly structured record of its own activity.
A significant part of the capability lies in how the components are connected. Persistent state, accessible context and feedback allow work to accumulate across tools and tasks. The consequences can reach beyond faster execution: an application can contribute to the process through which its own form is developed.
Saganode provides a concrete example. Brand knowledge from a marketing repository was structured into a Brand Universe using Codex. That context was then retrieved through MCP, the Model Context Protocol, and used in a design process involving the Vision Wizard, Codex, Figma and human decisions. Knowledge stored inside Saganode helped inform a new interface for Saganode itself.
The application indirectly participated in its own design.
From Outputs to Persistent Context
An isolated output has limited reach. A brief can guide a task, but its assumptions and decisions must often be explained again when the work moves elsewhere. Each transfer creates an opportunity to lose context or reconstruct it differently.
Structured, accessible knowledge changes that dynamic. Material created in one step remains available for subsequent steps. A model can retrieve it, develop it and return a revised structure to the same environment. Later actions begin from the state already reached.
The development of the Dragon Punk Universe illustrates this cumulative process. An initial topic structure provided the basis for iterative development in Saganode. Each addition changed the context available for the next step. As the Universe acquired structure, that structure could support further narrative development and analysis.
The useful unit of work therefore expands beyond the individual response. It includes the response, the state it changes and the possibilities that state makes available next.
Feedback Extends the Reach of Earlier Work
In a conventional sequence, information moves from source material to a finished artifact. Feedback introduces a return path: the artifact, or the knowledge used to produce it, can influence the conditions of future work.
The Brand Universe makes this visible. Knowledge originating in a marketing repository became structured context inside Saganode. Through Codex and Figma, it then informed decisions about the application that held it. The same material crossed from brand development into product design without requiring a complete manual reconstruction of its meaning.
The components contributed different functions. Saganode maintained structured knowledge. Codex interpreted and acted on accessible context. Figma provided the design environment. Human judgment determined direction, selection and acceptance. Their relationships allowed earlier work to become useful in a new setting.
This is a practical form of self-reference: an application supplies context for its own redesign. Describing it this way does not imply that the application has intentions or independently chooses its future. Its participation comes through the knowledge it preserves and makes available to the surrounding process.
A Research Lens on Coordination
Christoph Riedl’s paper, Emergent Coordination in Multi-Agent Language Models , offers a useful lens on the relationship between connection and collective capability. It studies multiple LLM agents in a guessing game, using an information-theoretic framework to examine higher-order structure.
The experiments distinguish temporal coupling from useful coordination. Distinct personas introduce stable differentiation; adding instructions to consider other agents’ behavior produces goal-directed complementarity. This suggests a question for connected workflows: how do differentiated contributions become coordinated around a shared task?
The Saganode example involves a heterogeneous arrangement of human judgment, a language model, persistent knowledge and a design tool. Riedl’s results do not establish emergence in this arrangement. They provide a research framework for examining how relationships between components contribute to collective behavior.
Which capabilities become possible because of the organization of those relationships?
Answering that question would require comparisons: what changes when persistent context is removed, when feedback is interrupted, or when responsibilities become less distinct? The practical example motivates the hypothesis; it does not settle it.
Direction Is Part of the System
As execution becomes easier, scope and selection become more consequential. A process capable of generating further structure still needs a basis for deciding which structure is useful, when to revise it and when to stop.
Human decisions are part of this dynamic. They determine the size of each iteration, evaluate whether a result supports the objective and redirect the process when expansion loses its purpose. The expansion problem in AI-assisted writing applies here too: continued production can create momentum without resolving what the work should become.
Persistent context can preserve earlier decisions, but those decisions remain open to review. Feedback becomes valuable when it supports correction as well as continuation. Otherwise, the system may accumulate assumptions as readily as useful knowledge.
Why the Change Becomes Hard to See
Once these relationships work reliably, expectations adjust quickly. A substantial redesign begins to feel like a small task alongside other work. The visible effort falls because much of the context has already been established and remains accessible.
That ease can obscure its own foundations. The final interface reveals little of the accumulated knowledge, decisions and connections that made it possible. Observers see a result; the working process depends on a history that is largely absent from the result itself.
The thesis is that persistent state, accessible context and iterative feedback enable capabilities that are best understood at the level of the whole working system. Earlier work continues to participate in later work. In the case of Saganode, that participation reaches back into the design of the application itself.
Source: Christoph Riedl, Emergent Coordination in Multi-Agent Language Models , ICLR 2026. arXiv version 4, revised April 28, 2026.