> ## Content Index > Fetch the complete content index at: https://www.elliotcsmith.com/llms.txt > Use this file to discover other available public pages before exploring further. # Deep Work and Agents - URL: https://www.elliotcsmith.com/deep-work-and-agents/ - Published: 2026-11-10T19:56:00.000Z - Updated: 2026-11-10T19:55:59.000Z - Author: Elliot Smith - Tags: AI I really like the ideas in [Deep Work](https://calnewport.com/deep-work-rules-for-focused-success-in-a-distracted-world/?ref=elliotcsmith.com). I've felt the difference between deep and shallow work and have, many times, tried to set goals to hit some amount of weekly deep work. The nature of work, especially in tech, has changed a lot this year and I noticed that depth doesn't come as naturally as it once did. For those who've not read Cal Newport's book the short of it is there is shallow and deep work. Shallow work is the work you need to attend but often feels more like administration than work. Deep work is, well, deep. You're in a state free of distraction and pushing your capabilities towards their limit. Newport argues that deep work is scarce and spending time there is both personally and professionally rewarding. Before we all had agents to contend with, deep work was typically achieved by blocking out some time in your calendar, picking a single task and concentrating on it for some meaningful length of time. The task chosen was as important as the distraction free environment, emails are not the kind of thing you can do with depth. The idea is that the time you spend on these deep problems is compounding during the focused session. Spending half an hour on something has a natural limit of how deeply you can work on it. By allowing longer periods of time the depth builds up and you can achieve outcomes that are not possible by someone trapped in the shallows. Now that more and more time is spent working with agents, I have started to wonder if they're a blocker to depth. Don't get me wrong, AI is a wonderful tool for leverage. With any of the recently released models you can achieve impressive things with far less time and effort. There's still a need to quality control but things that would have taken huge amounts of time and effort in the past are now close to trivial. That doesn't mean however that we're all now capable of deep work on command. Instead I would argue that what's happened is some things that were once only outcomes of deep work are now, in some sense, shallow. Further, they're now available to anyone with five minuets to spare and a basic ability in using AI. The other change I have noticed is that the way we interact with agents tends to be more like a meeting. You type some words, potentially clever ones, wait a while and then read some outputs. The way that agents work leads to natural breaks in the flow while you wait for them to crunch away some tokens. You may be very diligent and maintain your concentration while waiting for outputs but I think the loop of prompting and waiting means less overall compounding and therefore less depth. So what do we do here? One extreme might be to treat agents like email and silence them during attempts at deep work. I don't think this is the path to maximising returns on time but it would probably work to some degree. The other is to limit what kinds of work you do during that time. 'Working with AI' is combining a lot of things into one term in much the same way that you could use the term 'writing' to mean everything from posting on X to writing a novel. I've personally noticed that there are types of AI driven work that feel akin to AI-free deep work. In particular I feel it when spending time pulling apart ideas and challenging my thinking (similar to the very popular grill-me skill). While some people recommend less specificity and just letting AI make the judgement calls I am not convinced that deep work happens that way. It's probably fine if your goal is run of the mill but I still see a big return in spending time up front planning. The output seems to be better and I feel a deeper understanding of the problem which ultimately lets me go further with it. Another patterns I've seen and experienced is trying to swap depth for breadth. If there's wait time, start another thread. Try to maximise the number of threads such that you're never waiting on something. While I think this is potentially a path to more output I don't think that output is 'deeper' than it would be doing those tasks sequentially. On this, a warning. Part of the original deep work premise, and likely equally true today is that by default most professional environments are more incentivised towards total output than towards output depth. That's not to say they don't have depth but work places tend to have a low risk appetite for time spent. They'd rather a low risk median depth output than a high risk deep one. Unfortunately, there's no real metric or measure you can use to judge depth making this difficult to objectively assess. This means that oftentimes the measure is 'any output' rather than the quality of the output. As long as the output continues to appear there must be work going on. In any case I am still on the hunt for methods to find depth while working more with AI. I think the original premise still holds that if you can find a way to push past the average level of depth in your field your chances of success increase. There's also a great sense of personal satisfaction that comes from working deeply on things that I feel less and less as the herder of agent threads. If you've found a way to make this work, or disagree I'd be very interested to hear from you.