Context Window 96
I spent two days this week in a recording studio narrating the audio edition of my book for Penguin Random House Audio. I came out with a tired voice and an even bigger backlog of email and tasks than usual. As I read product announcements about delegation to personal AI agents, I wished they had been available to me. That will change quickly, and figuring out what this means for cooperative work is one of the big themes of this week’s newsletter.
It was OpenAI’s DevDay this week, and the company announced a new personal agentic AI platform, Dots. Like its direct competitor, Meta’s Muse, which I wrote about last week, Dots has a colourful, cartoonish, Pixar vibe. It certainly looks and sounds less threatening than, say, OpenClaw, and that speaks to the need to get individuals and companies to trust aspects of their working lives to an agent. Integration with Teams and Slack also helps with that.
Microsoft also made product announcements, upgrading Copilot with three capability areas: Home (including Chat and Cowork), Code and Autopilot. The latter is a personal agent previously known as Microsoft Scout, which will probably end up being the default for corporate employees who would otherwise have used Muse or Dots.
Short term, these personal agents are a potential productivity boost for early adopters. In the medium to long term, the question that interests me most is what happens when different agents start interacting with each other: for example, if someone in the rights department of a Big Five publisher gets their Autopilot agent to follow up with a licensee at a smaller house whose Dot picks up the query, or the author’s Muse enters the chat…
HBR published a good piece on this earlier in the month, focused on governance and data access, but I think there are a bunch of organisational behaviour factors to think through as we move to a mixed economy of H2A and A2A interactions. The first of these agents you encounter professionally may not be your own, and you’ll have to be ready to decide how you’ll work with it, or what information you’re prepared to give it. Figuring all this out is my early bet for one of the key themes of 2027.
In other OpenAI news, the company hired Sam Yam, co-founder of Patreon, to lead a new push on tools for creators. This is aimed more at the influencer end of the creative economy than writing and authorship, but it will be interesting to see how long-lived the idea is, and whether it results in features that would be useful for books. I suspect the issue is as much a cultural and PR one as a product challenge.
Relevant to this is PwC’s latest research on AI and jobs, published over the summer. This made a useful distinction between the trajectory for jobs which AI professionalises and those which it democratises. Professionalised jobs need more human input and expertise as a result of AI: for a commissioning editor, abundant content means that taste and judgement become more valuable. Democratised jobs are those where AI allows a non-expert to do something that previously required an expert. Which way OpenAI—or any other AI platform working with creatives—leans will be a critical question. Sorting your org chart or personal tasks into those categories could be an uncomfortable but useful early-warning exercise.
On the subject of creativity, HBR has a very relevant case study of workflows in a Dutch media company and what happens when AI collapses traditional creative processes and reviews. It makes some good practical recommendations, but, relating it to the agent stories above, my observation is that this describes problems with humans using AI in the workplace. If we add A2A interactions, it’s a whole extra layer of complexity.
Research from Newcastle University Business School and Hamburg University of Technology found that human-created work was valued more highly by consumers than AI-generated alternatives—but that among the youngest participants in the study, there was no value premium for human work. It’s one study, but worth asking whether a human value premium will hold for future generations of readers.
Digiday’s Publishing Summit last week saw a presentation from the New York Times on its approach to AI licensing, including its three core principles for evaluating potential deals: fair value exchange, control over content and partnership beyond a one-off payment. It would be unrealistic to think this would work for every publisher: the Times has a market position and bargaining power that smaller publishers do not, and for many, licensing will be a take-it-or-leave-it offer from AI companies. But it’s a useful aspirational framework.
Audible released new experimental AI features allowing listeners to explore the context of a book and even interact with characters. This is for a small number of Audible Original titles: rolling out something like this more broadly gets into the sort of questions about publisher and author approvals that Audible’s competitor Storytel is experiencing in some European markets. Two practical questions from this for every publisher: what do your distribution agreements with retailers say about AI use alongside your books, and if a retailer announces a feature like this, who communicates it to your authors?
What’s the best AI model for your use? That’s not just a question of Claude versus Copilot, but whether you select Fable, Opus or Sonnet for the task. There are lots of options, all of which claim to perform well on particular benchmarks. But, as this piece in Every highlights, some of those benchmarks are quite abstract, so it sets out how to build a personal evaluation framework—if you’re making decisions across a business or for key workflows, worth investing the time.
Finally, for a lighter take on the subject of model performance, a stop-press screenshot shared by my friend Tom Abba below. Pro tip: always read the vertical axis 😂