Tech Talk
Cowork By Michael E. Duffy
I know that it seems that all I talk heavily impacted by new tools like Claude Code from Anthropic and Codex from OpenAI, which aim to drastically reduce the need for human-generated code. And now Claude Cowork and OpenAI’s Frontier are poised to have a similar impact on non-technical work, using AI “agents” that automate tasks found outside software development: collecting data, analyzing it and creating reports or suggesting action. A couple of weeks ago, I attended a 40-minute webinar put on by Anthropic entitled “The Future of AI at Work: about these days is AI. But the rate at which changes are announced is unlike anything I’ve seen in the 50- ish years that I’ve been working in tech. Software development is being
to generate its own to-do list, showing how it planned to proceed. Although viewers never see exactly how long the process took, this time we got to see the outputs from Cowork, a spreadsheet and a PowerPoint which appear to be of professional quality. Of course, there is the question of whether the analysis is correct, and that is hard to judge from the demo. The issue is trust. If I asked a human coworker to perform this task, I would naturally check their work the first few times, to make sure that they hadn’t goofed up, especially since I would be presenting the data to others. In fact, I would probably ask them how they know that the analysis is correct. But assuming that they got it right, and could explain
Introducing Cowork” (you can watch it yourself at the following link: tinyurl.com/techtalk26-03; there are settings to watch at faster-than-normal speed). For me, it was eye-opening. The webinar started with an emphasis on the shift from conversing with chatbots to directing agents. What distinguishes agents from chatbots is the ability to use tools. For example, you can direct an agent to send an email or analyze the data in a folder. Claude Code already has this “agentic” ability, but it’s accessed via a command-line interface (which engineers are pretty comfortable with). But non-engineers saw the usefulness of Claude Code and started using it for other tasks, despite the less- than-friendly interface. Seeing that, Anthropic developed Cowork as a way to make it easier for non-engineers to benefit from directing agents to perform tasks. The first demo (at 21:39 in the webinar) started with the following request to Cowork: “Please help me find a good 2-bedroom apartment on Zillow. I’m looking in San Francisco around the Rincon Hill area (or South Beach/SOMA) under $6k a month. I want an in-unit washer/dryer and parking. Please create a spreadsheet comparing each listing’s link, address, rent, square footage, amenities and move-in date.” Cowork dutifully began assembling the information (although we didn’t see it run to conclusion). It’s easy to see how (assuming it worked correctly) Cowork could speed up an otherwise onerous data gathering task. The second demo (at 23:00) asked Cowork to analyze a folder containing weekly sales data and produce a spreadsheet allowing “what-if” modeling, as well as a PowerPoint presentation summarizing the data at a monthly business review. The prompt was pretty detailed, the sort of brief you might write for a competent coworker when asking them to do the same work. Much like that competent coworker, Cowork asked some clarifying questions and set to work. One of the things it did was
themselves, I would gradually come to trust their work. If you decide to use a tool like Cowork, you need to go through the same process of establishing trust in its work. In software development, we use tests to help ensure that code is correct, making it easier to trust code produced using AI. Other users need to develop their own checks (a task which AI can help with). When it was released as a “research preview” in January, Cowork was available to those subscribing to their $100 and $200 monthly plans. It is now available at the $20 per month level, although the amount you can do is more limited: Anthropic describes it as "perfect for quick tasks like organizing a folder or pulling together a short report.” Even at $2,400 a year for the “Max 20X” plan, it’s a bargain when compared to adding headcount. I was talking with my eldest daughter about AI. She works in marketing for a small architectural firm outside Washington, DC, putting together proposals for architectural work. This is exactly the sort of work agentic AI tools like Cowork are targeting. As I told her, it’s probably worth at least $20 a month out of her own pocket to learn how to use the tools and possibly demonstrate their effectiveness to her boss. It’s the same advice I would give to you. Agents are coming. They will change the way you do your job. Best to be ahead of the curve and invest in yourself. g
Michael E. Duffy is a 70-year-old senior software engineer for Electronic Arts. He lives in Sonoma County and has been writing about technology and business for NorthBay biz since 2001.
16 NorthBaybiz
Where Ideas Are Born 2026
Made with FlippingBook - Share PDF online