I heard that we often get cold/flu/sore throat when we get too cold outside, because the inside of our orifices is kept at a certain temperature to kill those bacteria/viruses. When we get too cold, we are unable to kill them fast enough, and get overrun. Staying in 70-100°C air for prolonged time must also heatshock those parts of our bodies, so I guess we kinda sterilize it that way.
I am running opus to make changes to my code then running the code. I am genuinely curious how we are having such disparate experiences here. And at this point, IMO you're in too deep not to share...
Genuinely wondering if you're running gastown or some other crazy mixture of agents pretending they're an AI startup. I get by with a developer agent and a reviewer agent ping ponging off each other encouraged to be rude, crude, and socially unacceptable about it.
Actually its just one opus aimed at a codebase with one goal, and instruction to spawn 2 subagents: one worker, that comes up with implementation plan, one validator that validates the proposed plan against my guardrails, and then return back to subagent1 to implement this, where the second subagent again tests the implementation.
One loop of this can take 20-60 min, and eat 2-5% of my week limit. I have to actively slow myself down to not burn trough more than 15-20% of my weekly limit in a day (as I also like to work on it on weekends)
Sadly I cant share the actual problem I am working on as its not my secret to disclose, but its nothing "crazy", and I am so surprised others dont have similar experience.
Very similar to what I am doing. How big is the codebase? My biggest was about 250K LOC and the usual is about 10K LOC. I am really curious about figuring this out because I'm genuinely puzzled.
My code base is two monorepos 10M+ lines. I have the same experience as you - run 3-6 agents with remote devcontainers and tmux and rarely break the 75% usage, never had the weekly limit stop me.
My observations are these things impact both quality and token consumption a lot.
- Architecture matters really- how messy code is and how poorly things are organized makes a big difference
- how context window is managed especially now with default 1M window.
- How many MCP servers are used. MCP burn a lot, CLI tools are easy , quicker and good ones don't even need any additional harness like skills etc, just prompt to suggest using them.
- Using the right tool matters a lot
- What can be done with traditional deterministic tools have to be done that way with agent controlling (or even building) the tool not doing the tool's work with tokens.
- for large refactors codemod tools AST parsers etc are better than having the agent parse and modify every module/file or inefficiently navigate the codebase with grep and sed.
- How much prep work/planning is put in before the agents starts writing code. Earlier corrections are cheaper than later after code is generated
Typically my starting prompts for the plan phase are 1-2 pages worth and take 30-60m to even put in the cli text box. With that, first I would generate detailed ADRs, documentation and breakdown issues in the ticketing system using the agent, review and correct the plan several times before attempting to have a single line written.
---
It is no different as we would do with a human, typing the lines of code was always easy part once you know what you want exactly.
It feels faster to bypass the thinking phase and offload to agent entirely to either stumble around and getting low res feedback or worse just wing it, either way is just adding a lot of debt and things slow down quickly .
When I got my Google AI Ultra, I could run it morning to evening at opus 4.6.
One month later I started hitting 5h limits when I was nearing my 5h window.
2 weeks after I hit my 5h limit 30 min into the morning. Cancelled my sub even quicker.
It doesn't, I have installed many Windows updates that didn't require a reboot. Even ones I expected to need an update, like an update to a graphics driver. Screen just went blank, then came back a second later.
AFAICT it's only updates to things that run at startup time that require a reboot, probably because NTFS doesn't allow you to write to a file that's currently opened (as opposed to nearly every Linux filesystem, which handles that just fine: the process that has the file opened continues to see the "old" file, while any that open it after the write will see the "new" file — but NTFS, probably due to internal architecture, can't handle that and so you have to reboot to change files that background services are using).
It has nothing to so with NTFS, but all with the Win32 API. The Windows kernel supports this file model, proven by WSL1. There is a blog post somewhere (Old New Thing?) stating the engineers would like to e.g. allow deleting a file even if there is still a program with with a file handle to it, but are concerned deviation from current behavior would cause more problems than it solves.
The reason that they want a reboot is that they do not want to support a system using two versions of the same library at the same time, let's say ntdll. So they would have to close any program using that library before programs that use the new version can be started. That is equivalent to a reboot.
And I completely understand the reason. For a long time when Firefox would update on Linux, the browser windows still open were broken; it opened resources meant for the updated Firefox with the processes runnung the non-updated Firefox. The Chrome developers mentioned [2] that the "proper" solution would be to open every file at start and pass that file descriptor to the subprocesses so all of them are using the same version of the file. Needless to say, resource usage would go up.
This isn't an NTFS thing. The I/O Manager implements NtLockFile. Applications can request exclusive byte-range write access to a file. And perhaps it is lazy programmers, or defaults, but they generally do.
I don't think Microsoft sees client machine reboots as an issue, and it used to be much worse when they used to be released weekly. On the server side, Microsoft expects that you'd implement some form of high availability.
NTFS on non-Windows follows the locking semantics of the underlying driver model/kernel, e.g. you can replace an in-use file on Linux. Likewise, using FAT on Windows you cannot replace an in-use file. This is just to demonstrate it isn't a file system-specific "issue" (if you feel it is one). It was a design decision by the original NT OS/2 development group.
Ultimately, the NT byte-range locking is a holdover from NT OS/2, where in OS/2 byte-range locking was mandatory.
There a enough apps that keep old files open, but also (re)open updated files that do not fit to the old, open ones, thus have all kind of issues.
(Subjectively Thunderbird has major issues with not restarting if libs it depends on get upgraded.)
I stopped answering support mails and tickets from users with long uptime with anything else than: reboot first.
And it was >>80% the cause of problems.
And yes, most times a logout would suffice, but with our users having >100d uptime with desktops and laptops, the occasional kernel update is done /en passant/ this way.
(The impatient could kexec and have the advantage of both. Or look at the output of "need restart" or "checkrestart". But I couldn't care less in case of end user devices)
Can‘t replace files that are in-use and that includes running programs or loaded DLLs. Linux can, it keeps the inode and only actually deletes upon termination of last access.
Ive read this many times, so I tried this a few times, giving it the benefit of the doubt, only to find the PC on login screen the following morning every time.
Ugh, I've had this happen over and over. I can't trust my laptop to actually shut down. I have to wait to see the light stay off for a couple seconds before I put it in my bag.
> Are Anthropic currently unable to sell subscriptions because they don’t have capacity?
Absolutely! Im currently paying $170 to google to use Opus in antigravity without limit in full agent mode, because I tried Anthropic $20 subscription and busted my limit within a single prompt. Im not gonna pay them $200 only to find out I hit the limit after 20 or even 50 prompts.
And after 2 more months my price is going to double to over $300, and I still have no intention of even trying the 20x Max plan, if its really just 20x more prompts than Pro.
At least my 2c why I think its helping
reply