TL;DR: Understanding the problem is still the bottleneck, surprise surprise
For some background, I've used a handful of different harnesses for development with AI Agents from Claude Code (my daily driver) to local agentic workflows with LM Studio and OpenClaw. Personally, it hasn't "multiplied" my output as many of the AI hype people claim and I think that's for one simple reason, I care about the output and how we got there to determine if we're solving the problem correctly and in a reasonable manner. Now, I can say that it allows me to form my thoughts in real English and for the most part will implement things in a similar fashion to how I would and reduces the number of iterations I go through to get there. Some features that I originally estimated at a week will now take a day or two, but I'm also doing more upfront work to research the problem and build context than I used to which isn't considered in that shortened time. It's likely that I'm able to build something of medium complexity in 50%-60% of the time that it used to take me, but I attribute a lot of that to being able to let the agent run while I'm in a call or otherwise occupied when that would previously be dead development time. A lot of my work tasks are things that are ambiguous or exploratory in nature and as such, don't have a lot in the way of information in the tickets. This way of working allows me to set the agent off on a task while I'm able to do research on the side or discuss with colleagues on the next problem in my list.
My typical flow will be to start off having the agent build a plan without doing any implementation. When using Claude Code I also enable verbose output so that I can read through the thinking output and potentially any commands that were run easily which is sometimes useful for debugging if things go wrong. I also use this as a learning opportunity, especially if I'm working with a new language or in a new-to-me part of the codebase. Once that plan is written up, I'll review that, make changes if necessary, then let the agent go to town. Just like with then planning step, I'll review the thinking output and if I'm watching it as it's running I can steer the agent one direction or the other if necessary. Finally, once the agent is done, I review the output. It's taken a while, but I'm finally close to having the output tuned to be similar to what my team likes to see from a style standpoint so there's not much iterating there anymore. Usually my iterations are on doing things like method extraction or taking the opportunity to do some refactoring for readability and maintainability.
Outside of work, I really haven't seen anywhere where my life can be improved by agentic workflows. I live a fairly boring life on purpose and don't need agents bothering me throughout the day, I have enough app notifications and emails as it is.
I find it hard to give agents permission to freely go through my system, emails, text messages and such. When using them for development I restrict them to the working directory and don't allow them to pull in random dependencies from the internet, only whatever is out there on Nuget, npm, etc. I also have guidance in my system prompts to ask before doing anything if there's any amount of ambiguity in the action versus my ask and to not delete files or directories if they're not tracked by some type of version control. I've also yet to come across a need to connect an agent directly to a running application or database, instead providing schemas and example data if necessary for the task.
I do think that there's some room for me to build out some skills to help handle my typical support tickets using read-only connections so that might be something I look into next. I have already written a skill that will help with code reviews, primarily because there's a lot of areas of the codebase that I'm still not familiar with even a year and a half into my current job. I find that this has helped tremendously because it'll catch some things that I might miss because of issues going end-to-end while the logic of an individual file looks okay. I have found that there's a fairly high false-positive rate in the skill but I also don't want to spend all of my credit on that each day so if nothing else it gets me to take a second look at some things.
Wrapping up, I'm finding my usage of AI and AI Agents to definitely be a positive on my work, but I'm not sure it's a multiplier, at least in the work that I typically do with the restrictions I put on them. If I wanted to go hog wild and let it have access to everything, yeah maybe, but that's not for me right now. I'm probably preaching to the choir, but my main problem is still understanding the work and the desired approach to solving it. Agentic development has definitely helped my on the implementation side of the house because I can let it go and write the code for me, but I still provide the context and steer the ship. The agent is my driver and I'm the backseat driver.