For former ICs who aren’t used to watching out for overwork and burnout as new managers of agents, there’s decades of research studying why people overwork once they’re responsible for (and occasional blockers for) workers that remains just as applicable whether they’re human or bot. Start with this HBR article and explore their Overwork keyword in more depth, perhaps: https://hbr.org/2024/01/want-to-be-a-better-leader-stop-thin...
I just implemented a task queue in my hard fork of paseo; so when I got all the ideas but need to wait for implementation on local models, I just set them all up in the queue and the agents go to work. They start in a single agent and work synchonously because thats how most tasks are, and if the agent has a question, the next item gets shunted to a new session and it continues. Just tested it out on a walk and it's surprisingly effective. Then once it's exhausted, it notifies me (TBD) and I can go in and resolve the remaining issues.
It's definitely an issue of how people feel obligated to feed the beast, but some of it is just how the tools are setup. They can definitely be improved so you don't feel as if youre slowing down because it's idling an figment of a feature or whatever.
But then, I'm just working a small niche non-software driven org and my goal is a "total awareness" kit that will be the hub for anyone who comes after me to carry on the required engineering and science of my company.
well, I spent a few months wandering the desert of offerings.
Once Qwen3.8-Flash-Next demonstrated it could consistently implement Test Driven Development, I decided to hard fork the "best" work flow that made sense to me, which was paseo.
Then once opencode decided they wanted to rewrite opencode2, I've also started a hard fork of opencode 1.
Who also abandoned that for some other larger goal.
I've also installed openviking to test out some of the memory stuff.
But to the point: local modals & hardware (all hardware in fact) has a sweet spot of performace. The longer the context grows, the slower they get. Both with restarting any session and just getting pre-fill and token gen.
So, dynamic context prun compresses the context and it's typically compressing it at the top of the stack. So to preserve KV cache, it's putting compressed tools+messages on top. But that eventually is exceeded. There's only so much compressed infomation you can get to; and those sessions eventually acrue too much.
So then I added an eviction command for the agent or I to trigger. Eviction gates at a ratio, 5:1 on my setup, so when the context get to the max limit, the agent is instructed to start eviction. This obviously differents from compression, as we're now digging deeper into the stack. But what it's doing is adding waypoints instead of compression.
A waypoint allows the coherence of the context to remain without all the context pollution that comes from the compression command.
So zooming out, we set a min/max context of say 64k/128k. The model moves through min/max space doing regular compression. Find things, summarize things, progress. But eventually it hits >128k, so the system tells it to evict, and evict is at the bottom of the kv cache, so there is a cost at that point. But that cost is paid for in the speed. The evict's goal is to hit a good ratio 5:1, and to get beneath the 64k. So the agent is aware of what I call context pollution, as context grows, pollution grows.
It took awhile to get all the language together, the agent's prompt is very thin compared to what opencode provides. We treat the agent prompt as a starting point and not some swiss army knife of anything. We provide some skills, but not much. Most is tools.
Anyway, eventually the hard fork of paseo will directly have all this codified; but the rolling window of context is what's valuable and allows us to create implementation plans and ocmplete them. I know this works because I wouldn't hard fork active libraries if I couldn't keep this ball moving.
And I hate react, but that's what paseo's rendering is doing and it's quite capable of adding ui features.
it's actually very valid these days - we can take any book about team leadership / management and apply it to managing AI agents, it will work most of the times
I have definitely found work sprawling out across a larger span of hours than it used to occupy, but I don't feel that I'm actually putting in more effort. I'm just working more intermittently, with long breaks throughout the day while the robots grind away at the tasks I've set for them. For decades I was strict about confining work to work hours, but these days it doesn't feel like a burden to check in on work in the evening, since I have plenty of time during the day for housework and personal matters.
i completely believe this. i've been working from around 7AM to somewhere between 10-midnight seven days a week since june. (from home, trying to get my LLC going, but _still_)
I strongly recommend you stop doing this in the new year, find more support, and also that you spend more time than you might initially plan to on catching up with people over the festive period.
Two reasons:
1) burnout is real, and it can also devastate your ability to focus and enjoy really anything; you won’t know it until you get there but it can make you ill or make you more vulnerable to getting ill. I have spent two years trying not to believe I have actual dementia, and it is only now I am beginning to recover. I think agentic AI use risks a kind of burnout that is a terrible mix of something akin to doomscrolling and professional overwhelm. It’s for sure possible to burn out without it — I did — but my feeling as I get my life together and confront AI is that some people are already struggling with this; it’s paradoxically maybe only my burned out brain that stops me from getting that bad.
2) when you disappear from people’s lives a part of you disappears with it. Six months is quite a long time; don’t let it go longer than that. It is enough to strain personal connections you might miss in the future.
Good luck with your business, but take care of yourself :-)
as a person who went through burnout twice (even before AI) I'm warning you to be careful, it always happens unexpectedly and you realize "shit I'm done" when it's a bit late
proper rest without any work-related app in front of you (ideally outside or with other people) heals pretty fast, so make sure to clear your head that way
Cynical take: A form of selection bias, already unproductive deadweight employees may feel no pressure to use AI. Longer workdays relfect an already established distribution.
throw in the forced RTO commute time to sit at an open office desk, commanding the agents while others teammates do the same around you, and now thats what I call collaborative productivity maximalization for a measly 3 more hours a day.
I can't shake the feeling that RTO and mandated agent herding and vague threats of imminent human obsolescence are the manager class' way of bullying the previously free laptop class into a regular wage slaves.
it feels a lot more feasible to fit in one more small task in at the end of the day since you’re a lot more productive which leads to me working a bit longer
For me, I'm in that mode right now because Fable 5.2 just dropped. As a brand-new frontier release, it will never be as fast or as smart as it is now (and I DGAF who says otherwise). So I'm working like the protagonist in "Flowers for Algernon," feeling his IQ slipping away and wanting to make the most of it while he has it.
Probably unhealthy, but dammit, it's just too much fun.
For former ICs who aren’t used to watching out for overwork and burnout as new managers of agents, there’s decades of research studying why people overwork once they’re responsible for (and occasional blockers for) workers that remains just as applicable whether they’re human or bot. Start with this HBR article and explore their Overwork keyword in more depth, perhaps: https://hbr.org/2024/01/want-to-be-a-better-leader-stop-thin...
I just implemented a task queue in my hard fork of paseo; so when I got all the ideas but need to wait for implementation on local models, I just set them all up in the queue and the agents go to work. They start in a single agent and work synchonously because thats how most tasks are, and if the agent has a question, the next item gets shunted to a new session and it continues. Just tested it out on a walk and it's surprisingly effective. Then once it's exhausted, it notifies me (TBD) and I can go in and resolve the remaining issues.
It's definitely an issue of how people feel obligated to feed the beast, but some of it is just how the tools are setup. They can definitely be improved so you don't feel as if youre slowing down because it's idling an figment of a feature or whatever.
But then, I'm just working a small niche non-software driven org and my goal is a "total awareness" kit that will be the hub for anyone who comes after me to carry on the required engineering and science of my company.
Would be interested to hear more about how you have this set up logically
well, I spent a few months wandering the desert of offerings.
Once Qwen3.8-Flash-Next demonstrated it could consistently implement Test Driven Development, I decided to hard fork the "best" work flow that made sense to me, which was paseo.
Then once opencode decided they wanted to rewrite opencode2, I've also started a hard fork of opencode 1.
I pulled in the opencode plugin from: https://github.com/Tarquinen/opencode-dynamic-context-prunin...
Who also abandoned that for some other larger goal.
I've also installed openviking to test out some of the memory stuff.
But to the point: local modals & hardware (all hardware in fact) has a sweet spot of performace. The longer the context grows, the slower they get. Both with restarting any session and just getting pre-fill and token gen.
So, dynamic context prun compresses the context and it's typically compressing it at the top of the stack. So to preserve KV cache, it's putting compressed tools+messages on top. But that eventually is exceeded. There's only so much compressed infomation you can get to; and those sessions eventually acrue too much.
So then I added an eviction command for the agent or I to trigger. Eviction gates at a ratio, 5:1 on my setup, so when the context get to the max limit, the agent is instructed to start eviction. This obviously differents from compression, as we're now digging deeper into the stack. But what it's doing is adding waypoints instead of compression.
A waypoint allows the coherence of the context to remain without all the context pollution that comes from the compression command.
So zooming out, we set a min/max context of say 64k/128k. The model moves through min/max space doing regular compression. Find things, summarize things, progress. But eventually it hits >128k, so the system tells it to evict, and evict is at the bottom of the kv cache, so there is a cost at that point. But that cost is paid for in the speed. The evict's goal is to hit a good ratio 5:1, and to get beneath the 64k. So the agent is aware of what I call context pollution, as context grows, pollution grows.
It took awhile to get all the language together, the agent's prompt is very thin compared to what opencode provides. We treat the agent prompt as a starting point and not some swiss army knife of anything. We provide some skills, but not much. Most is tools.
Anyway, eventually the hard fork of paseo will directly have all this codified; but the rolling window of context is what's valuable and allows us to create implementation plans and ocmplete them. I know this works because I wouldn't hard fork active libraries if I couldn't keep this ball moving.
And I hate react, but that's what paseo's rendering is doing and it's quite capable of adding ui features.
hey, is your fork source available anywhere?
it's actually very valid these days - we can take any book about team leadership / management and apply it to managing AI agents, it will work most of the times
I have definitely found work sprawling out across a larger span of hours than it used to occupy, but I don't feel that I'm actually putting in more effort. I'm just working more intermittently, with long breaks throughout the day while the robots grind away at the tasks I've set for them. For decades I was strict about confining work to work hours, but these days it doesn't feel like a burden to check in on work in the evening, since I have plenty of time during the day for housework and personal matters.
Non-AI users have been fired and their work shifted to AI users.
That isn't showing up in the unemployment figures.
The admin seems to fire bureaucrats that deliver bad numbers. Are there any third party validation of employment numbers?
i completely believe this. i've been working from around 7AM to somewhere between 10-midnight seven days a week since june. (from home, trying to get my LLC going, but _still_)
I strongly recommend you stop doing this in the new year, find more support, and also that you spend more time than you might initially plan to on catching up with people over the festive period.
Two reasons:
1) burnout is real, and it can also devastate your ability to focus and enjoy really anything; you won’t know it until you get there but it can make you ill or make you more vulnerable to getting ill. I have spent two years trying not to believe I have actual dementia, and it is only now I am beginning to recover. I think agentic AI use risks a kind of burnout that is a terrible mix of something akin to doomscrolling and professional overwhelm. It’s for sure possible to burn out without it — I did — but my feeling as I get my life together and confront AI is that some people are already struggling with this; it’s paradoxically maybe only my burned out brain that stops me from getting that bad.
2) when you disappear from people’s lives a part of you disappears with it. Six months is quite a long time; don’t let it go longer than that. It is enough to strain personal connections you might miss in the future.
Good luck with your business, but take care of yourself :-)
i can tell you have a kind soul
thank you for your response
https://www.youtube.com/watch?v=WWIopTjcYbc
as a person who went through burnout twice (even before AI) I'm warning you to be careful, it always happens unexpectedly and you realize "shit I'm done" when it's a bit late
proper rest without any work-related app in front of you (ideally outside or with other people) heals pretty fast, so make sure to clear your head that way
But like legitimately, what is it that you are doing, that takes up that much time?
If a computer is a bicycle for the mind, AI is an e-bike for the mind. It's simply easier to put in longer sessions.
For suckers who arent paid overtime
Cynical take: A form of selection bias, already unproductive deadweight employees may feel no pressure to use AI. Longer workdays relfect an already established distribution.
That is not cynical take. That is wish fulfillment take, that assumes that a.) world is fair b.) worthy people are kind of like me.
It's all because of "just one more prompt" and then waiting for an hour to prepare everything for the over-night run :D
This reminds me of when I wrote Spark jobs.
throw in the forced RTO commute time to sit at an open office desk, commanding the agents while others teammates do the same around you, and now thats what I call collaborative productivity maximalization for a measly 3 more hours a day.
I can't shake the feeling that RTO and mandated agent herding and vague threats of imminent human obsolescence are the manager class' way of bullying the previously free laptop class into a regular wage slaves.
Two times 5 hour shifts. Oh, no usage anymore today? Let's call it a day.
it feels a lot more feasible to fit in one more small task in at the end of the day since you’re a lot more productive which leads to me working a bit longer
For me, I'm in that mode right now because Fable 5.2 just dropped. As a brand-new frontier release, it will never be as fast or as smart as it is now (and I DGAF who says otherwise). So I'm working like the protagonist in "Flowers for Algernon," feeling his IQ slipping away and wanting to make the most of it while he has it.
Probably unhealthy, but dammit, it's just too much fun.
[flagged]