## Origins, Echos & Empires Electric Dreams started with curiosity. ![[Breaking Lines Poster.jpeg|200]] When ChatGPT appeared at the end of 2022, our first encounters with it were not particularly strategic. We asked it questions, played with it, and were surprised by the complexity of the responses. Quite quickly, though, we became interested in the gap between what we were experiencing through actually using these tools and some of the conversations happening around us. A lot of the early education conversation was understandably about risk. Would students use it to cheat? How would we detect that? What should the rules be? What position should an institution take? Those are legitimate questions, but they did not feel like the whole conversation. We work in an art and design context, so we were also interested in a different question: what happens when these technologies meet a discipline? What happens when they meet photography, fashion design, performance, illustration or assessment? And importantly, what happens when you put them into the hands of practitioners and ask them to explore rather than simply respond to a predetermined position? That became one of the central ideas of Electric Dreams. So rather than beginning with AI and asking teaching teams to become experts in AI, we begin with disciplinary expertise. The proposition is quite simple. Our teaching teams already understand their disciplines. They understand their histories, their practices, their materials and their students. Our job is to create a space in which they can examine where technological change intersects with that knowledge. That is why the project has always been collective. We do not think these questions are particularly useful when answered by individuals in isolation. We are interested in what happens when a teaching team talks together, experiments together, and gradually begins to articulate a collective position. And one of the things we have learned through the project is that the technology itself can sometimes help us do that. We have used recording and transcription to capture conversations. We have used generative tools to help teams return to what they have said. And we have begun experimenting with agents as another layer in that process. The important thing for us is that this is not about automating teaching. We are interested in whether a machine can take on some of the processing, organising and pattern-finding so that more time is available for the human conversation. That has become increasingly important to the project: individual agency _and_ collective agency. ## Agents & Agency That brings us to agents. And there is a connection here with some of our earlier work. A couple of years ago, alongside Nadia Malik and Ben Turnbull we wrote a creative piece called [_Future Echoes: A Prologue and a Tale_](https://www.tandfonline.com/doi/full/10.1080/13569783.2025.2506698). We worked with Anthropic’s Claude throughout the process and, eventually, began describing Claude as our dramaturg. At the time, we did that through repeated prompting. We would tell it what we were trying to do, correct it, give it more context, ask it to change tone, and gradually build up a working relationship inside one conversation. Looking back - today we would probably formalise much of that through an agent. Microsofts use of ‘Agents’ is annoying. Agents in this case are not autonomous, but more like ‘projects’ in other systems. Instead of starting every interaction from scratch, we can define a role and give the agent a set of instructions. So we can think about three layers: Copilot gives us the environment. The agent gives us a particular role or purpose. And the instructions help determine how that agent behaves. The instructions might tell it what Electric Dreams is, the language we use and what we value. They can also tell it what not to do. For example, we can tell it not to treat AI as the starting point of every conversation. We can tell it to begin with disciplinary practice. We can tell it to distinguish between speculation, evidence and recommendation. And we can tell it to avoid presenting generated material as though it represents a collective position unless that position is actually evidenced. That is the interesting part for us. The intelligence is not simply in the model – it sits in how carefully we construct the conditions around the model. An agent without useful context is still just a general-purpose generative system. An agent with a clear role, good instructions and relevant project knowledge starts to become something much more specific and meaningful. Not an expert replacing the people in the room, but a tool configured to work within a particular methodology. We are asking: “What role do we want this particular agent to perform within this particular human process?” ## Conclusion So if we reduce all of that to three things, the first is that **instructions matter**. What you ask a system to pay attention to significantly shapes what it gives back. Second, **context matters**. The more clearly we can situate the agent within a real disciplinary, pedagogic or organisational context, the more useful the interaction becomes. And third, **human judgement always remains central**. For Electric Dreams, the interesting possibility here is not replacing the conversation. It is whether tools like these can help us retain, organise and return a collective conversation to the people who produced it. The journey for us has moved from asking a chatbot a question, to working collaboratively with generative AI, to thinking much more deliberately about how we design the conditions in which an agent responds.