Context Window 79
There are two big themes this week: what new AI agents are making possible, and the need for a more nuanced discussion of AI’s impact on authors and publishers. In particular, Matt Webb’s first-person reflection on agents will resonate with any publisher shipping products internationally in 2026.
HBR published its third annual survey of how AI is actually being used, compiled by Marc Zao-Sanders. It’s a must-read, with several strong themes around the limits of AI, the sorts of boundaries we need if we are to use it effectively, and the use of AI agents in the workplace. But the line that really stuck with me was on organisational impact: ”We see AI mostly being used to achieve modest, uncontroversial wins… and few [use cases] where business processes are being fundamentally rethought.” This chimes very strongly with my experience. It’s worth remembering that a modest, uncontroversial win is still a win. And enough of them can be really significant if there is a systematic approach to aggregating marginal gains. The use of AI agents in the workplace was the highest new entry in the HBR survey. For many publishers, that means Claude Code and Cowork. The best practical resource I came across this week was this Anthropic training video on how to effectively prompt agentic systems. The specifics are about Claude, but the underlying principles could be applied to other agentic platforms. Claude’s lead in agentic systems hasn’t gone unnoticed by major competitors such as Microsoft, which this week announced a new personal work agent, Scout, which works with Microsoft 365 apps and is accessed through Teams. It is based on the open-source OpenClaw technology, which has mixed reviews for reliability (an AI security researcher described it ignoring her instructions and deleting all of her emails), though Microsoft claims its model has professional-grade controls. Similarly, Meta launched a new Business Agent product which allows companies to interact with customers across Facebook, Instagram, WhatsApp and Threads. It also allows integration with third-party tools like Shopify and Zendesk. Many publishers rely on those platforms for marketing and sales, so it’s potentially highly relevant, but beware of even more lock-in to Meta platforms: it is advertised as free to start, but paid subscriptions are on the way. Meta may need to reassure potential customers on security after Meta’s AI assistant helpfully reset recovery emails for high-profile Instagram accounts. Multi-factor authentication for everything, people. If this discussion of agents is making you nervous or unsure of what this might actually mean for you and your work, my friend Matt Webb wrote a nice, first-person piece about why he’s been reluctant to embrace personal agents and what changed his mind (semi-spoiler: it’s an experience that any publisher shipping globally in 2026 might share). The Competition and Markets Authority, the UK’s competition regulator, imposed new conduct requirements on Google Search, including the ability for publishers to opt out of their content being used in AI search summaries, and requirements for proper attribution and links. While the main beneficiaries of this will be big news and information publishers, the CMA’s conduct requirements define publishers broadly as “all parties that make content available on the web”, so book and journal content (I am thinking particularly of Open Access material published online) will fall into the scope of the requirements. There’s an interesting data point in the reporting of the Scottish Book Trade Conference: Edinburgh University Press CEO Nicola Ramsey confirmed that 80% of EUP’s authors have opted into licensing of content to AI platforms. It’s a useful reminder that author opinion on AI is nuanced by type of publishing, career stage and other factors. The impact of AI on early-career employment is another area that is more nuanced than is often suggested. Researchers at the New York Federal Reserve and the London School of Economics published complementary analysis arguing respectively that remote work rather than AI explains much of the rise in graduate unemployment, and that studies which blamed AI for reduced early-career hiring may have confused the impact of AI and remote work. The key point here is not that there’s a green light to use AI without consideration of career impact, or even that people should be back in the office more—it is that early-career work is as much an apprenticeship as it is task performance. AI needs to be used without pulling up the ladder that the next generation of publishers need to climb. Practically, given that it’s often experienced staff who get the best productivity improvements from AI, could that time saved be deliberately reinvested in mentoring and career development for junior colleagues? Finally, I’ve written about a few of The Economist’s experiments with AI in recent weeks, but here’s an example of a personal experiment by one of their journalists. Outgoing defence editor Shashank Joshi built an interactive simulator to help readers understand the UK’s notoriously Byzantine defence budget, using Claude Opus 4.8 and ChatGPT. That may or may not be of interest to you in terms of subject matter—the more interesting thing for publishers and content developers is that Shashank confirmed in a Bluesky thread that the first draft took just 5-10 minutes to produce and a couple of hours to refine. It illustrates vividly that most of our mental models for the time and money to create interactive media are collapsing in real time.