Writing
Agents belong inside the tools people already use
Copilot in Word, pyReality in Jupyter, a MIDI controller wired into Rekordbox: the same decision, made three times.
When I describe my work on Copilot in Word, the phrase I keep using is "all within the familiar Word environment". It sounds like a detail. I think it is the decision that mattered most.
I was among the first designers to bring transformer-based language models into Microsoft’s ecosystem, mainly Word, and the work helps people draft, summarize, rewrite and edit documents with conversational agents. But I do not believe the model is what made those capabilities usable by millions of people. It was that nothing about the person’s workflow had to change. They were in Word, with their document, their formatting and their habits, and the agent came to them.
That is the pattern I have come back to for years, and it predates my AI work.
pyReality was my master’s dissertation at the University of Edinburgh. It is a Python library that lets you build mixed reality data visualizations directly inside a Jupyter notebook, view them through WebXR on a head-mounted display, and never leave the notebook. The problem it solved was fragmentation. Traditional tools required separate platforms for modelling, rendering and viewing, so anyone who wanted to explore data in three dimensions had to switch between several tools, and none of them fit the analysis workflow they already had. Putting the whole thing into the notebook, alongside the Python code the analyst was already writing, was the entire design. Early testers singled out how accessible and fast it was for prototyping. That speed came from the absence of a context switch, not from any clever feature.
Makerarm went the same way. It was a desktop fabrication system that could 3D print, laser cut, carve, plot and assemble. I led the development of Makerarm.io, a browser-based control interface with real-time 3D visualization and a cloud-based architecture that allowed control from any device. Makers already had a browser. Meeting them there was part of what made a multi-process fabrication system feel approachable rather than industrial.
Even my own MIDI controller followed the rule. I built it because my two-channel DJ controller limited me to two tracks, and adding a third meant switching layers and breaking the flow of a set. The answer was a custom Arduino device that plugged straight into Rekordbox, the software I already performed with. I later mapped it to Ableton for studio use. The point was never the hardware. It was that the new capability arrived inside the setup I already knew.
Why this matters more for agents
First, trust. An agent asks for a lot of it. It will act on your document, your data or your machine. Trust is much easier to extend inside a place you already understand, where the surrounding tools behave as they always have and you can see the result in context. Moving the same capability into a new app resets that trust to zero.
Second, context. An agent is only as good as what it can see. Inside Word the agent has the document. Inside a notebook, pyReality has the data the analyst has already loaded. Inside Rekordbox, every knob on the controller maps to a control that already exists. A standalone agent has to be told all of that, and every step of telling it is friction the person notices.
Third, adoption. Nobody is looking for a new tool. They are looking to finish what they are doing. A capability that appears in the place they already are gets used. One that requires a new tab, a new account and a new mental model competes with everything else on their desk.
There is a cost. Designing inside an established tool means working within its conventions, its constraints and its history, and that is harder than starting with a blank canvas. But those constraints are the reason the work lands. A former colleague once said that I use technical constraints as creative opportunities rather than limitations. I would put it more plainly: the tool people already use is the brief. Start there.
More in the writing section, or see the design process these pieces come from.