- It is divided into by default 2048 regions so it will be setting region size to 4mb means with an 8gb heap, and that you’ll get an inefficient collection once you have a 2mb+ object to deal with. So there is a value to setting a bigger heap size to get a bigger region size. The max is 32mb region size then after that iirc the region count bumps. https://docs.oracle.com/en/java/javase/17/gctuning/garbage-f...
- The young gen size also something that resizes dynamically and so having a bigger max heap will let that get larger and avoid pressure and promotion of objects that would otherwise just never get promoted in the first place (stop the world to move from new to old gen). I’ve never seen this stated anywhere I can think of but believe that I’ve observed it many thousands times.
- The reason that a Full GC occurs is because the application allocates too many objects that can't be reclaimed quickly enough, this isn’t just something that is always happening in the background , I would call full gcs a tuning problem, not everyone agrees in a given org fwiw, but I’ve been able to tune them out for nearly all workloads. https://docs.oracle.com/en/java/javase/17/gctuning/garbage-f...
TLDR , as long as I have enough ram for the working set (page cache etc) I’ve had good luck cranking g1gc very high in prod and avoiding long collection times all together. You see other large deployments do similar (was big in the Cassandra space back in the day, Instagram off the top of my head was running very large heaps with g1gc before anyone else I knew of).
No, not especially, the client (Zed in this case) can add some things (like @ing a file or class) and hint stuff but in all seriousness the harness is doing the things I’ve mentioned
You can see it from the different transcripts from the different tooling harnesses (´agents’ in agent client protocol speak) you can also just read the source for pi or zed or anything here don’t take my word for it.
You will see different web search from pi and codex, different tool calls, different context compaction strategies, etc.
The client has some control over this stuff but very little, like zed can pass in mcp servers to use or files to add to the context but the tooling harnesses can ignore those requests
25 years experience, work at an AI startup building AI dev tools (tooling harness, review bots, etc), I use lots of different techniques all the time to test our products and competitors products out.
Fable is at once amazing and awful. I can see how having it build websites would be awesome.. building anything I’ve needed some precision in functionality it has been a constant battle of it plausibly building something then on substantial manual digging (like the review bots always miss it) I will find that one of the fundamental features is all smoke and mirrors.
To be fair all models can and will do this (especially anthropic) but Fable takes the cake because it builds such impressive UX and you can manually test the feature out and it « works » then you will find days later one of the features violated one of your constraints in a devilishly fiendish way.. that is not at all what you want or can accept. Fable generated work already holds my record for the most reverted commits.
To be clear it’s also solved several features I thought I was going to have to give up on and hand code as GPT-5.5 and Opus-4.x we’re failing miserably.
I would only reach for it for nasty corner cases that everything else sucks at.
Final point, it is the king of UX work so far, not even close.
I hate to say it as I think the author is well intentioned but I think this is every bit as silly as the perspectives it is shooting down
1. this is prone to taking the fringe perspectives and making them "everyone"
2. most of the points are highly highly interpretive (growth could be pressure, laziness, FOMO, self deception, real valid growth.. we don't know yet)
LLMs are definitely the most transformative thing I have seen in 25 years in tech, but I still think like every other hype cycle there is a lot of lying and self delusion, I don't really think any of us, neither myself nor the author really know what we have here.
Harness does a lot of things potentially, but this depends on the harness.
1. Manages what is loaded in the context (I.e. sent to the model), this can be about compaction, summarization, removing things it judges no longer relevant, proving memory
2. Handling retries and how to resend data , lot of model specific behavior here
3. Wires up tools like web search or lsp servers
4. Skills and how to use them
5. Plugins/sdks
6. Management of rules (AGENTS.md)
About everything interesting really, a ton of work goes into this and getting it right is surprisingly hard
I’m sure you’re very intelligent and capable so I suspect we work in different problem spaces if you have not seen this, but I definitely think the responses are at times very very junior and I find myself having to explain first principles. Fable less so, but Opus routinely will make very naive assumptions about retry logic, protocols it supposedly has training material on, and it will very often miss the forest for the trees.
This isn’t exactly saying how stupid anyone is but I’d definitely have been concerned about a human’s reasoning ability and understanding of logic if they’d given me similar answers.
- It is divided into by default 2048 regions so it will be setting region size to 4mb means with an 8gb heap, and that you’ll get an inefficient collection once you have a 2mb+ object to deal with. So there is a value to setting a bigger heap size to get a bigger region size. The max is 32mb region size then after that iirc the region count bumps. https://docs.oracle.com/en/java/javase/17/gctuning/garbage-f...
- The young gen size also something that resizes dynamically and so having a bigger max heap will let that get larger and avoid pressure and promotion of objects that would otherwise just never get promoted in the first place (stop the world to move from new to old gen). I’ve never seen this stated anywhere I can think of but believe that I’ve observed it many thousands times.
- The reason that a Full GC occurs is because the application allocates too many objects that can't be reclaimed quickly enough, this isn’t just something that is always happening in the background , I would call full gcs a tuning problem, not everyone agrees in a given org fwiw, but I’ve been able to tune them out for nearly all workloads. https://docs.oracle.com/en/java/javase/17/gctuning/garbage-f...
TLDR , as long as I have enough ram for the working set (page cache etc) I’ve had good luck cranking g1gc very high in prod and avoiding long collection times all together. You see other large deployments do similar (was big in the Cassandra space back in the day, Instagram off the top of my head was running very large heaps with g1gc before anyone else I knew of).