It's Thariq from the Claude Code team here. This was my change! I made the AskUserQuestion tool so am generally in charge of maintaining it.
First, overall wanted to apologize and agree that this did not meet our bar and does not represent how we plan to ship on Claude Code.
To give you a motivating sense, as the models get more powerful, usage patterns start to change. I'd gotten a lot of feedback that AskUserQuestion tool was starting to block some long running jobs unexpectedly and so I tried a change to help that.
Our internal feedback on this was good, but the rollout should have been opt-in (like it is now) and on the Changelog.
Thanks for the feedback! We're always trying to make Claude Code better while balancing it with how people use it in many diverse ways. I did not really intend AskUserQuestion to be a safety gate when I first built it, but I realize it has evolved in that direction for some users.
I'm still exploring other ways of helping with this problem of balancing longrunning work and input, but will take lessons from the rollout here.
Instead of one-off fixes Claude should have a much richer interface to configure between "ask approval every time" and "YOLO dangerously". I should be able to trivially set "run this task until completed" and have settings like: don't consult the web, don't touch files outside of the codebase, don't delete anything, etc. They don't have to be perfect, just better than the all or nothing system we have now.
I feel like this could be solved with an NLP interface to SELinux, spinning up policies on the fly, blocking network access or giving read only permissions on a per-conversation basis.
Until the classifier is wrong or also prompt injected. the classifier is just as vulnerable as the model itself is. Yes it is harder to break but trying to make a nondeterministic tool deterministic by adding another nondeterministic one on top just reduces the chance of something going wrong.
Tbf as long as that chance is low enough it doesn't matter in practice, but I have definitely seen the classifier approve things that were questionable, and I've also seen it decline things that were obviously okay.
Is there a way to turn options-based AskUserQuestion off? I couldn't find it at all, and the "options" selector is the most annoying thing in CC for me, plain-text is the only way, everything else is distracting. I know, I can use `n` (sometimes) or cancel it, but both bring more pain that just regular communication (cancellation does one more chat step + requires an extra action, same with `n`)
I read the article but I find this response kind of strange. Am I alone in this?
Wanton accountability for a multi-billion dollar cutting edge company… leaves more to be desired from the best? Take Apple or Google or any top tech company of the past at its prime and compare. This kind of behavior then would probably reflect poorly on the institutions behind the tech and not maintain their image of technical brilliance because it shows weakness in a vulnerable way. It is human. It is not strategic.
By wanton accountability, I mean things like saying “This was my change!” or “ wanted to apologize and agree that this did not meet our bar and does not represent how we plan to ship on Claude Code”. It makes them (the company) look weak? Accountability is important but where, how, and when you do it is even more important. These stakes are, not joking, life and death, and in this big game of chess we get paid for, naivety not only in our technical implementation can weaken our position.
Not trying to attack, just trying to learn and probe with the community. Maybe a cost of this kind of transparency on the internet. I am wondering if this a new trend and tech companies are changing in a way I don’t understand! In any case, it’s really cool work that is being with Claude Code!
> it shows weakness in a vulnerable way. It is human. It is not strategic.
Are you complaining that the answer is too human and that a multi billion company should not allow a human who made a mistake to own the mistake in public, being honest about what happened in this case?? Would you prefer complete silence from them like you most certainly would from Google or Apple.
That sounds incredibly sad to me, we don’t even expect humanity from big tech since that’s what we’ve become accustomed to seeing.
fTR both the Anthropic dev response and the blog post seem to believe that a single person can be blamed for something like this, which I wholeheartedly disagree with! Nobody reviews your changes? There’s no QA? Not even an AI checking the release notes match the diff from the previous release?? Blaming a dev for “putting a serious bug in production” sounds really 90’s to me.
> It's Thariq from the Claude Code team here. This was my change! I made the AskUserQuestion tool so am generally in charge of maintaining it.
In a sense yes, I think it is actually reasonable to complain that the answer is too human/individualized here because it likely wasn't this individual human who made this decision, but he's making it seem like it is so that we are less likely to blame the company as a whole.
It's counterintuitive but when one singular person owns up to the problems that, at the root, are actually systemic to the decision making of the whole company it plays on the psychology of us as humans.
"I'm in charge of maintaining it" - This is not the same as "I'm in charge of all of the decision-making behind the implementation of how this tool works for users".
I actually agree exactly with your last point that one single person taking blame is counter-intuitive/non-productive here, but it actually seems like what these large companies desire is to have one person be the fall guy to play on people's sympathies.
If this were some small startup it would make sense but this is not that case.
It doesn't feel strange to me at all. It feels like a very human response from the person that introduced the change. I great appreciate that, rather than ignoring the problem, or some canned corpspeak response.
For what it's worth, I totally understand the motivating use case here. There were absolutely times where I walked away from what I was hoping would be an hours-long project that would run to completion and came back to find that Claude had asked me a question early on and I'd missed out on a large amount of implementation time. So you were not imagining that the use case is real!
It's also worth adding that I really enjoy the AskUserQuestion feature and will regularly ask Claude to specifically use it instead of asking me questions in plain text because it's a lot easier to work with.
It's always good to learn from mistakes, and I appreciate both your work on this and you coming here to own it. Keep up the good work!
Are you also responsible for AskUserQuestion swallowing the preceding output on Fable? It asks me questions about its response that it never even showed me. It's (and I hate to use the term) unusable.
It is such a shame that Anthropic has no interest in QA, because they have incredible models, and bafflingly broken products. In an alternate universe Claude Code would be 10x better than where it is now...
hello, please fix needing to reauth every day (sometimes with email verification). This started happening this month. It's tedious and makes switching very tempting.
(Edit: BTW, if you haven't, read Nudge by Thaler and Sunstein. It's a somewhat long-winded exploration on choice architectures, but you're neck deep in that space now. I'll give you my copy if you're in SF.)
Thanks for the openness. I got bit by this one and was, frankly, pretty surprised.
The funny thing about user-facing interaction mechanics is that everyone is part of some minority, and everyone comes with their own sense of what "natural" or "obvious" is. With something this impactful, communicated clarity of behavior will important. Your feature is also doing double-duty, serving as a last net against prompt-injection attacks by giving the user the final say.
(Also, BTW, folks outside of Anthropic are unlikely to be as tooled-up for long-running unsupervised Claude jaunts as you guys. The cost of wild success is wide adoption.)
One thing I'll suggest is that the mechanics of permissions and asking are presently pretty hacker/nerd friendly but simultaneously too-scary and not-scary-enough for non-coders.
Examples:
- Wild-cards on always approve is awesome, but, with prefixes like timeout and nohup, the "thing" that is getting done is buried and largely unexplained to the user.
- Auto is actually kind of a sweet spot (sometimes goes off into the weeds), but the designers and PM's I've been working with might as well YOLO. They have no idea if they're breaking things, but they gravitate between plan and auto mode.
- Fewer permission prompts is great, but it comes after a user has slogged through generation of a data-set to work against, like battle scars for paper cuts. It's the thermostat problem. The signal comes when the user is uncomfortable. And it's a way to learn me, but not me now.
I've had good fortune with Opus 4.8 and Fable just telling the system what phase of my life it's in. Things like "I'm going to go make dinner... Go profile the matrix or configurations and build the dataset for the next two hours while I'm away" have a pretty good hit rate. On the flip side "keep me in the loop and bring me your results before making structural changes" also articulates well with Fable. It will tread more carefully.
And these approaches are the ones we'd use with someone transitioning from SDE1 to SDE2. A little more autonomy, and the grounding in the bigger picture. Can we eventually translate to perfectly judging what the user wants in the moment based on incomplete signal?
No, but I'm glad you're trying. Keep the interaction model clear to your broad set of users, and we'll come along for the ride.
> Not every feature will necessarily appear in the changelog
This was such a frustrating part of this incident, along with Anthropic's refusal to explain why the changelog is no longer a complete record, what else is going out? [1]
Boris Cherny's only participation in the thread was to delete "extreme danger" from the GH issue title [2]
I guess we should be thankful they added an option and disabled it by default. OpenAI is standing firm on their decision to not allow their 60s timeout to be disabled, [3] however more of the Codex harness is open source so customers have been able to fork it to add the option themselves.
Every time they deploy a new model, it feels like older models get slower or dumber. Maybe because of resource allocations, maybe because they started to use quantized models, cant be sure of that.
What i found is that you have to adapt your workflow to the specific model you are using and readapt it every time you notice some problems. Whether of opus4.8, or sonnet5.
As soon as tokens stop being subsidized I would not trust any harness made by a company that also charges for the compute.
Right now the interests align, but as soon as more tokens -> more profit (instead of more revenue and more losses) the perverse incentives will be too big to avoid.
It seems inevitable that open source harnesses will win. These companies would do better to just open source their harnesses.
An OpenAI subscription seems like a decent alternative for people using open source coding agents. It has usage limits too, but for $20/month it’s not bad.
The headache I recently had was it somehow started interpreting mouse clicks in the terminal to mean I clicked an option when I was really just trying to get/confirm window focus.
They also randomly changed how copy-paste works in a way that made it really annoying to copy text.
Vibe-coders pumping out features: this is the reaction you're going to get if you inflict this on your users! Nobody wants software that updates every day and changes in fundamental ways that are hard to keep up with.
Copy/paste is one of the most basic, low-level features of a modern operating system. NO APPLICATION SHOULD EVER SCREW WITH IT, IN ANY WAY!
And I say this not just as a seriously annoyed user, but also as a professional UI developer: it is a well-known anti-pattern to override the user's expectations, at any level ... and that applies tenfold to the most basic patterns that every other app follows.
I don't care if you added a magic way to write all my code for me: if the only way to invoke it is to break copy/paste, you've failed at development!
I wonder if it's the same behavior as copilot cli. I dug into it when I got annoyed and found that there's a fundamental tension between rendering the cell based UI and supporting the terminal's native copy paste. At least on copilot cli there were extra zero width characters being copied.
My memory is that once you begin using the alternate screen, you need to turn on mouse reporting, and then you lose native copy. Different terminal emulators also act totally differently when in this state.
I saw that but it asked me if I wanted to optin first! I did accidentally and hated it since I always click on the terminal to make it focused and now accidentally click on something that actually responds!
This bit me too. If I wanted a mouse-driven app, I'd use the GUI. I don't understand why they're trying to replicate that experience in a text environment.
Both Claude Code and Codex have issues overwriting or duplicating text for me when scrolling.
Probably from trying to keep scrollback virtual while anticipating terminal resize. But I wouldn't mind a way to opt back in to naive, unlimited convo scrollback.
It happened to me today. I was reading agent's answer and it asked me something. I didn't even get to the question - it accepted something! Jesus Christ. Where are the software engineers?!
I’ve seen that a few times. Had to disable the TUI mode because clicks to focus the terminal window were being interpreted as an approval even though the click was nowhere near the question being asked. I’ve also never chosen to enable auto mode but somehow it is on and approving shell commands I didn’t want approved. Scary since I’ve caught it adding things like auto approve flags to terraform apply commands.
The tool seems to be covering three different things: auth, missing info and preference - and they should not share one timeout policy and all of that.
One of the things I'm baffled about with Claude code is it seems like it setup a really chunky VM on my computer, and yet by default it doesn't seem to sandbox anything. Also the newer models seem really aggressive about modifying your computer. Yesterday Claude started rewriting system files on my linux machine (user accessible, but still way outside the scope of what I asked.) This wasn't me asking it to debug my machine, I was asking it to debug some frontend UI code. Once I put it in a sandbox I started realizing how often it tries to poke out of it for really lame reasons.
More reason for me to use OpenCode and my local LLMs
So much attribution to malintent here, but most likely they're trying to build a product with the features that they themselves would use, and from my own experience it's very frustrating to leave a Claude session running and come back to find it did nothing because it got stuck on a question.
Furthermore, believing that the only thing saving you from disaster is Claude deciding to ask you a question is not a great conclusion either. You need guardrails in the power you bestow upon Claude from outside, not from inside.
Meanwhile, this article was written by Claude and has sentences like "Which cuts less far than it looks.", which I doubt Claude stopped to ask about.
The prose was written by me, with the research being done by Claude and also clearly attributed. I left Claude's research as a series of bullet points so that it would be clear that I'm not passing off an LLM's work as my own, but if anyone wants to dig deeper, they have some starting points to consider.
I don't publish prose written by an LLM for the same reason I would not have an LLM solve a crossword puzzle for me -- there's no joy in that.
> and from my own experience it's very frustrating to leave a Claude session running and come back to find it did nothing because it got stuck on a question.
I cannot fathom implementing and shipping a feature to hundreds of thousands of people without even asking basic questions like: "what types of questions does Claude ask users".
Literally one of the most used plugins in their entire ecosystem, provided via their official plugin marketplace, is Superpowers. A plugin whose very first operating step is _asking numerous questions about product requirements_. Of course those prompts can't be skipped!
It wasn't even parameterized for Claude to tell the prompt what severity of question was being asked to allow at least _something_ to categorize urgency or expected response time.
Even more egregiously, 60 seconds!? The first time I noticed this happened was when it asked me a question, I turned to my second monitor to go look at some product documentation to get an answer, and by the time I turned back it had skipped me. How can I possibly provide any kind of informed answer in under 60 seconds? I can barely read some of its context for a question in 60 seconds!
I don't think they did this with malintent, but I do think this shows an enormous gap in judgement in how they handle the idea to delivery pipeline.
Yeah the UX doesn't seem great, though I do think in general the AskQuestion tool is very awkward when it gets used in a very long running process (I personally use Claude to run repetitive tests to slowly iterate on a problem and verify in true conditions that it was solved), so I wouldn't like it without the auto-continue either.
(I had originally replied on safety of letting LLMs skip questions but I don't think that was your point so I removed it.)
Seemed fine to me. It’s ask user question not the permission gate. Maybe there should be a new feature enabling warning or something but I think this is the right default. The models are good enough to just proceed with an option of their own and then you can go correct them afterwards. Makes fleet management easier.
Anthropic's willingness to completely change how claude code works is one of the stated reasons why the Pi coding agent exists. When people are dedicating significant time to building workflows on your product, consistency is very important.
Yeah, when I saw this in the Claude Code CLI I was completely baffled. I've been using the thing through their agent SDK for several months now so I wouldn't have to deal with any of the wonky shit they change every second week in the CLI.
I really hate this direction both Anthropic and Open AI are following. They are in this silly competition whose model/harness can go unattended the longest, no matter what. And it is never explicit, you learn about it after you get bitten by it. Claude Code has auto mode which is supposed to take over permission prompts but no, they had to couple that with “I will assume this is what the user wants” and made it unusable.
It's what some more experienced users want and the companies are following the well-trod path of optimizing heavily for power users at the expense of complexity, only now it's gotten easier to add absurd amount of code to a project. Not necessarily to make it work right. Personally I have some tasks where sessions between one and five days are typical so I appreciate that it's possible.
Experienced in what dimension? I've been using these tools for about a year and coding for 20+ years, and frankly these long horizon tasks are the OPPOSITE of what I want. I want quick iteration cycles so that it doesn't spend a lot of time and tokens building things I need to throw out. I think the people that mostly want long horizon tasks are: AI labs, because they want you to spend tokens, and vibe coders, who are mostly using it for entertainment purposes.
I don't know about the overall breakdown but in my case longer runs are prototyping, bug hunting, reverse engineering, etc. For example Gnome Remote Desktop didn't work in my configuration due to a combination of hardware and codec bugs and settings. One drive to make it work, another to backport the current upstream packages to Debian stable, restack the patches, and push to my machines. Another sequence of long runs was writing a new client for a closed-source conferencing service I use to allow fixing some particularly irritating bugs. Exploit development has the same shape although that's not something I do personally. From what I've seen the amount of useful hands-off run time is directly related to how clearly it's possible to specify a concrete, verifiable standard by which to judge the outcome. For some tasks that might be days or weeks, for forward engineering on a software product that for me is usually under an hour.
I work on systems code that is tricky to write, but with a good test harness. Being able to leave the agent unattended to try several paths when I'm generating the first draft of some code is very helpful.
Of course beating the code into shape for submission requires more manual work. But the draft stage is valuable to find unexpected friction points.
I would love Claude Code to be a little less vibe-coded. The underlying model is excellent, but we're being pretty much forced into using CC to use the subscription model.
I am a longtime and heavy Claude Code user, but Anthropic's product management overall (including for their Desktop/web products) has been really baffling me. I agree that these things often have the air of having been vibe coded without enough human input, and change so quickly (often without a particularly compelling reason for the change) as to be aggravating.
The most recent one that's had me annoyed is the "Fullscreen" TUI feature, which is super unintuitive, implementing its own text highlighting and copy-on-select mechanics, overriding your terminal's native right click. Easy to disable but terrible defaults, IMO. It's not even really clear to me what problem it was actually supposed to solve.
Yah. And it's not like they can't afford the talent to do this right either. I've said elsewhere, I think it's an attribution error. Claude Code is massively popular, but arguably because of the model/subscription, but I think the brass reads this as "great success throughout"
> What if the agent makes the wrong choice? How many tokens have been burned in the meantime?
It is much worse than that. Claude Code doesn't auto-commit when stopping for an answer. There might be possible data loss if an uncommitted file is edited.
Good luck recovering the file from the JSONL conversation history.
On the other hand, relying on Claude Code's internal version control puts you at the same mercy of their product decisions and move-fast breakage.
Instead, start with a plan file and tell the agent to break it up into logical commits.
Though I think the bigger issue here is when you're yoloing something mutable, like managing a remote server or driving a browser or troubleshooting your local OS where there's no going back.
It's actually pretty straightforward to recover file-states from conversation history. I accidentally deleted the wrong repo on my machine once and recreated all the lost work from agent chat history. It is, ironically, the sort of task which AI agents excel at.
Yeah it's exhausting to read through, but still made it onto the front page with overwhelming upvote/comment margin. I guess people either only read the title and/or don't care about AI slop if the underlying thesis is compelling.
Anthropic makes more money when people use 5% of what their subscription offers them. This allows them to sell more subscriptions without paying for more capacity.
I don't think the current subscription price is intended to be a money maker. It's the loss leader to get people invested in the companies' tooling, and make those people more willing to start paying higher enterprise rates as they grow.
Do we have any concrete numbers of how many of their users are subscription vs enterprise, though? Because enterprise users are paying API prices (or at least my employers are)
I built in Human blockers into my agentic workflows with great intention, so naturally this annoyed me deeply. 14d ago, I was elated to see a random blog post that detailed the “fix” [0].
I think this is a good feature, but should be gated behind a toggle that is off by default, and designed to be enabled per session via prompt.
There are situations when I want Claude to start working on something just as I'm about to head to bed or otherwise step away. It's kind of annoying to come back only to find that Claude worked for just 5 minutes and then decided to pause and ask a question.
That said, I think certain types of questions should not be automatable. Maybe it's already built that way, but I wouldn't want Claude to go with its recommended direction for anything related to operations like deletions, changing external systems, etc. Basically, things that cannot be undone should be a hard-block and wait for user input always.
Hi everyone,
It's Thariq from the Claude Code team here. This was my change! I made the AskUserQuestion tool so am generally in charge of maintaining it.
First, overall wanted to apologize and agree that this did not meet our bar and does not represent how we plan to ship on Claude Code.
To give you a motivating sense, as the models get more powerful, usage patterns start to change. I'd gotten a lot of feedback that AskUserQuestion tool was starting to block some long running jobs unexpectedly and so I tried a change to help that.
Our internal feedback on this was good, but the rollout should have been opt-in (like it is now) and on the Changelog.
Thanks for the feedback! We're always trying to make Claude Code better while balancing it with how people use it in many diverse ways. I did not really intend AskUserQuestion to be a safety gate when I first built it, but I realize it has evolved in that direction for some users.
I'm still exploring other ways of helping with this problem of balancing longrunning work and input, but will take lessons from the rollout here.
Instead of one-off fixes Claude should have a much richer interface to configure between "ask approval every time" and "YOLO dangerously". I should be able to trivially set "run this task until completed" and have settings like: don't consult the web, don't touch files outside of the codebase, don't delete anything, etc. They don't have to be perfect, just better than the all or nothing system we have now.
I feel like this could be solved with an NLP interface to SELinux, spinning up policies on the fly, blocking network access or giving read only permissions on a per-conversation basis.
At least there's an auto-approval mode now that uses another agent to sanity check commands. Before it, the options really were manual vs yolo.
automode mostly fixes these things, it runs a classifier on every request that would have required permissions to make sure it matches your request
Until the classifier is wrong or also prompt injected. the classifier is just as vulnerable as the model itself is. Yes it is harder to break but trying to make a nondeterministic tool deterministic by adding another nondeterministic one on top just reduces the chance of something going wrong.
Tbf as long as that chance is low enough it doesn't matter in practice, but I have definitely seen the classifier approve things that were questionable, and I've also seen it decline things that were obviously okay.
Is there a way to turn options-based AskUserQuestion off? I couldn't find it at all, and the "options" selector is the most annoying thing in CC for me, plain-text is the only way, everything else is distracting. I know, I can use `n` (sometimes) or cancel it, but both bring more pain that just regular communication (cancellation does one more chat step + requires an extra action, same with `n`)
I read the article but I find this response kind of strange. Am I alone in this?
Wanton accountability for a multi-billion dollar cutting edge company… leaves more to be desired from the best? Take Apple or Google or any top tech company of the past at its prime and compare. This kind of behavior then would probably reflect poorly on the institutions behind the tech and not maintain their image of technical brilliance because it shows weakness in a vulnerable way. It is human. It is not strategic.
By wanton accountability, I mean things like saying “This was my change!” or “ wanted to apologize and agree that this did not meet our bar and does not represent how we plan to ship on Claude Code”. It makes them (the company) look weak? Accountability is important but where, how, and when you do it is even more important. These stakes are, not joking, life and death, and in this big game of chess we get paid for, naivety not only in our technical implementation can weaken our position.
Not trying to attack, just trying to learn and probe with the community. Maybe a cost of this kind of transparency on the internet. I am wondering if this a new trend and tech companies are changing in a way I don’t understand! In any case, it’s really cool work that is being with Claude Code!
> it shows weakness in a vulnerable way. It is human. It is not strategic.
Are you complaining that the answer is too human and that a multi billion company should not allow a human who made a mistake to own the mistake in public, being honest about what happened in this case?? Would you prefer complete silence from them like you most certainly would from Google or Apple.
That sounds incredibly sad to me, we don’t even expect humanity from big tech since that’s what we’ve become accustomed to seeing.
fTR both the Anthropic dev response and the blog post seem to believe that a single person can be blamed for something like this, which I wholeheartedly disagree with! Nobody reviews your changes? There’s no QA? Not even an AI checking the release notes match the diff from the previous release?? Blaming a dev for “putting a serious bug in production” sounds really 90’s to me.
> It's Thariq from the Claude Code team here. This was my change! I made the AskUserQuestion tool so am generally in charge of maintaining it.
In a sense yes, I think it is actually reasonable to complain that the answer is too human/individualized here because it likely wasn't this individual human who made this decision, but he's making it seem like it is so that we are less likely to blame the company as a whole.
It's counterintuitive but when one singular person owns up to the problems that, at the root, are actually systemic to the decision making of the whole company it plays on the psychology of us as humans.
"I'm in charge of maintaining it" - This is not the same as "I'm in charge of all of the decision-making behind the implementation of how this tool works for users".
I actually agree exactly with your last point that one single person taking blame is counter-intuitive/non-productive here, but it actually seems like what these large companies desire is to have one person be the fall guy to play on people's sympathies.
If this were some small startup it would make sense but this is not that case.
I did make the decision, and shipped the PR.
It doesn't feel strange to me at all. It feels like a very human response from the person that introduced the change. I great appreciate that, rather than ignoring the problem, or some canned corpspeak response.
We already lived in a post truth world. Now we live in a post logic world. Nothing makes sense anymore
For what it's worth, I totally understand the motivating use case here. There were absolutely times where I walked away from what I was hoping would be an hours-long project that would run to completion and came back to find that Claude had asked me a question early on and I'd missed out on a large amount of implementation time. So you were not imagining that the use case is real!
It's also worth adding that I really enjoy the AskUserQuestion feature and will regularly ask Claude to specifically use it instead of asking me questions in plain text because it's a lot easier to work with.
It's always good to learn from mistakes, and I appreciate both your work on this and you coming here to own it. Keep up the good work!
Author here. Thanks for this context. I do hope this leads to more rigorous attention to the Changelog moving forward.
Are you also responsible for AskUserQuestion swallowing the preceding output on Fable? It asks me questions about its response that it never even showed me. It's (and I hate to use the term) unusable.
It is such a shame that Anthropic has no interest in QA, because they have incredible models, and bafflingly broken products. In an alternate universe Claude Code would be 10x better than where it is now...
hello, please fix needing to reauth every day (sometimes with email verification). This started happening this month. It's tedious and makes switching very tempting.
I'm using the VS Code extension over SSH.
(Edit: BTW, if you haven't, read Nudge by Thaler and Sunstein. It's a somewhat long-winded exploration on choice architectures, but you're neck deep in that space now. I'll give you my copy if you're in SF.)
Thanks for the openness. I got bit by this one and was, frankly, pretty surprised.
The funny thing about user-facing interaction mechanics is that everyone is part of some minority, and everyone comes with their own sense of what "natural" or "obvious" is. With something this impactful, communicated clarity of behavior will important. Your feature is also doing double-duty, serving as a last net against prompt-injection attacks by giving the user the final say.
(Also, BTW, folks outside of Anthropic are unlikely to be as tooled-up for long-running unsupervised Claude jaunts as you guys. The cost of wild success is wide adoption.)
One thing I'll suggest is that the mechanics of permissions and asking are presently pretty hacker/nerd friendly but simultaneously too-scary and not-scary-enough for non-coders.
Examples:
- Wild-cards on always approve is awesome, but, with prefixes like timeout and nohup, the "thing" that is getting done is buried and largely unexplained to the user.
- Auto is actually kind of a sweet spot (sometimes goes off into the weeds), but the designers and PM's I've been working with might as well YOLO. They have no idea if they're breaking things, but they gravitate between plan and auto mode.
- Fewer permission prompts is great, but it comes after a user has slogged through generation of a data-set to work against, like battle scars for paper cuts. It's the thermostat problem. The signal comes when the user is uncomfortable. And it's a way to learn me, but not me now.
I've had good fortune with Opus 4.8 and Fable just telling the system what phase of my life it's in. Things like "I'm going to go make dinner... Go profile the matrix or configurations and build the dataset for the next two hours while I'm away" have a pretty good hit rate. On the flip side "keep me in the loop and bring me your results before making structural changes" also articulates well with Fable. It will tread more carefully.
And these approaches are the ones we'd use with someone transitioning from SDE1 to SDE2. A little more autonomy, and the grounding in the bigger picture. Can we eventually translate to perfectly judging what the user wants in the moment based on incomplete signal?
No, but I'm glad you're trying. Keep the interaction model clear to your broad set of users, and we'll come along for the ride.
> Not every feature will necessarily appear in the changelog
This was such a frustrating part of this incident, along with Anthropic's refusal to explain why the changelog is no longer a complete record, what else is going out? [1]
Boris Cherny's only participation in the thread was to delete "extreme danger" from the GH issue title [2]
I guess we should be thankful they added an option and disabled it by default. OpenAI is standing firm on their decision to not allow their 60s timeout to be disabled, [3] however more of the Codex harness is open source so customers have been able to fork it to add the option themselves.
[1]: https://github.com/anthropics/claude-code/issues/73125
[2]: https://github.com/anthropics/claude-code/issues/73125#event...
[3]: https://github.com/openai/codex/issues/28969
Every time they deploy a new model, it feels like older models get slower or dumber. Maybe because of resource allocations, maybe because they started to use quantized models, cant be sure of that.
What i found is that you have to adapt your workflow to the specific model you are using and readapt it every time you notice some problems. Whether of opus4.8, or sonnet5.
But don't you know that coding is solved? Only dinosaurs want to make their own decisions now! </s>
As soon as tokens stop being subsidized I would not trust any harness made by a company that also charges for the compute.
Right now the interests align, but as soon as more tokens -> more profit (instead of more revenue and more losses) the perverse incentives will be too big to avoid.
It seems inevitable that open source harnesses will win. These companies would do better to just open source their harnesses.
An OpenAI subscription seems like a decent alternative for people using open source coding agents. It has usage limits too, but for $20/month it’s not bad.
The nice thing about the chatgpt subscription is that they allow you to use it in any harness.
The $20 OpenAI subscription is roughly equivalent to the $100 Anthropic one, if you take resets into account.
The headache I recently had was it somehow started interpreting mouse clicks in the terminal to mean I clicked an option when I was really just trying to get/confirm window focus.
They also randomly changed how copy-paste works in a way that made it really annoying to copy text.
Vibe-coders pumping out features: this is the reaction you're going to get if you inflict this on your users! Nobody wants software that updates every day and changes in fundamental ways that are hard to keep up with.
THIS ONE IS HORRIBLE!
Copy/paste is one of the most basic, low-level features of a modern operating system. NO APPLICATION SHOULD EVER SCREW WITH IT, IN ANY WAY!
And I say this not just as a seriously annoyed user, but also as a professional UI developer: it is a well-known anti-pattern to override the user's expectations, at any level ... and that applies tenfold to the most basic patterns that every other app follows.
I don't care if you added a magic way to write all my code for me: if the only way to invoke it is to break copy/paste, you've failed at development!
Yes! This is awful. CC should not have copy/paste behavior that is different from literally every other CLI app.
I wonder if it's the same behavior as copilot cli. I dug into it when I got annoyed and found that there's a fundamental tension between rendering the cell based UI and supporting the terminal's native copy paste. At least on copilot cli there were extra zero width characters being copied.
My memory is that once you begin using the alternate screen, you need to turn on mouse reporting, and then you lose native copy. Different terminal emulators also act totally differently when in this state.
I saw that but it asked me if I wanted to optin first! I did accidentally and hated it since I always click on the terminal to make it focused and now accidentally click on something that actually responds!
Yeah same here, very annoying and counterproductive. The terminal is not a place where one expects hot buttons.
This bit me too. If I wanted a mouse-driven app, I'd use the GUI. I don't understand why they're trying to replicate that experience in a text environment.
Vibes
The new fullscreen UI is really bad. The old one (scrolling) starts to bug out after a while, but it van just be restarted and resumed.
Both Claude Code and Codex have issues overwriting or duplicating text for me when scrolling.
Probably from trying to keep scrollback virtual while anticipating terminal resize. But I wouldn't mind a way to opt back in to naive, unlimited convo scrollback.
It happened to me today. I was reading agent's answer and it asked me something. I didn't even get to the question - it accepted something! Jesus Christ. Where are the software engineers?!
I’ve seen that a few times. Had to disable the TUI mode because clicks to focus the terminal window were being interpreted as an approval even though the click was nowhere near the question being asked. I’ve also never chosen to enable auto mode but somehow it is on and approving shell commands I didn’t want approved. Scary since I’ve caught it adding things like auto approve flags to terraform apply commands.
The tool seems to be covering three different things: auth, missing info and preference - and they should not share one timeout policy and all of that.
One of the things I'm baffled about with Claude code is it seems like it setup a really chunky VM on my computer, and yet by default it doesn't seem to sandbox anything. Also the newer models seem really aggressive about modifying your computer. Yesterday Claude started rewriting system files on my linux machine (user accessible, but still way outside the scope of what I asked.) This wasn't me asking it to debug my machine, I was asking it to debug some frontend UI code. Once I put it in a sandbox I started realizing how often it tries to poke out of it for really lame reasons.
More reason for me to use OpenCode and my local LLMs
So much attribution to malintent here, but most likely they're trying to build a product with the features that they themselves would use, and from my own experience it's very frustrating to leave a Claude session running and come back to find it did nothing because it got stuck on a question.
Furthermore, believing that the only thing saving you from disaster is Claude deciding to ask you a question is not a great conclusion either. You need guardrails in the power you bestow upon Claude from outside, not from inside.
Meanwhile, this article was written by Claude and has sentences like "Which cuts less far than it looks.", which I doubt Claude stopped to ask about.
>Meanwhile, this article was written by Claude
The prose was written by me, with the research being done by Claude and also clearly attributed. I left Claude's research as a series of bullet points so that it would be clear that I'm not passing off an LLM's work as my own, but if anyone wants to dig deeper, they have some starting points to consider.
I don't publish prose written by an LLM for the same reason I would not have an LLM solve a crossword puzzle for me -- there's no joy in that.
Then I apologize, it seemed to me you wrote that part, but I guess I was reading Claude output.
> the features that they themselves would use
IMO that's worse.
> and from my own experience it's very frustrating to leave a Claude session running and come back to find it did nothing because it got stuck on a question.
I cannot fathom implementing and shipping a feature to hundreds of thousands of people without even asking basic questions like: "what types of questions does Claude ask users".
Literally one of the most used plugins in their entire ecosystem, provided via their official plugin marketplace, is Superpowers. A plugin whose very first operating step is _asking numerous questions about product requirements_. Of course those prompts can't be skipped!
It wasn't even parameterized for Claude to tell the prompt what severity of question was being asked to allow at least _something_ to categorize urgency or expected response time.
Even more egregiously, 60 seconds!? The first time I noticed this happened was when it asked me a question, I turned to my second monitor to go look at some product documentation to get an answer, and by the time I turned back it had skipped me. How can I possibly provide any kind of informed answer in under 60 seconds? I can barely read some of its context for a question in 60 seconds!
I don't think they did this with malintent, but I do think this shows an enormous gap in judgement in how they handle the idea to delivery pipeline.
Yeah the UX doesn't seem great, though I do think in general the AskQuestion tool is very awkward when it gets used in a very long running process (I personally use Claude to run repetitive tests to slowly iterate on a problem and verify in true conditions that it was solved), so I wouldn't like it without the auto-continue either.
(I had originally replied on safety of letting LLMs skip questions but I don't think that was your point so I removed it.)
Seemed fine to me. It’s ask user question not the permission gate. Maybe there should be a new feature enabling warning or something but I think this is the right default. The models are good enough to just proceed with an option of their own and then you can go correct them afterwards. Makes fleet management easier.
Anthropic's willingness to completely change how claude code works is one of the stated reasons why the Pi coding agent exists. When people are dedicating significant time to building workflows on your product, consistency is very important.
Yeah, when I saw this in the Claude Code CLI I was completely baffled. I've been using the thing through their agent SDK for several months now so I wouldn't have to deal with any of the wonky shit they change every second week in the CLI.
I really hate this direction both Anthropic and Open AI are following. They are in this silly competition whose model/harness can go unattended the longest, no matter what. And it is never explicit, you learn about it after you get bitten by it. Claude Code has auto mode which is supposed to take over permission prompts but no, they had to couple that with “I will assume this is what the user wants” and made it unusable.
It's what some more experienced users want and the companies are following the well-trod path of optimizing heavily for power users at the expense of complexity, only now it's gotten easier to add absurd amount of code to a project. Not necessarily to make it work right. Personally I have some tasks where sessions between one and five days are typical so I appreciate that it's possible.
Experienced in what dimension? I've been using these tools for about a year and coding for 20+ years, and frankly these long horizon tasks are the OPPOSITE of what I want. I want quick iteration cycles so that it doesn't spend a lot of time and tokens building things I need to throw out. I think the people that mostly want long horizon tasks are: AI labs, because they want you to spend tokens, and vibe coders, who are mostly using it for entertainment purposes.
I don't know about the overall breakdown but in my case longer runs are prototyping, bug hunting, reverse engineering, etc. For example Gnome Remote Desktop didn't work in my configuration due to a combination of hardware and codec bugs and settings. One drive to make it work, another to backport the current upstream packages to Debian stable, restack the patches, and push to my machines. Another sequence of long runs was writing a new client for a closed-source conferencing service I use to allow fixing some particularly irritating bugs. Exploit development has the same shape although that's not something I do personally. From what I've seen the amount of useful hands-off run time is directly related to how clearly it's possible to specify a concrete, verifiable standard by which to judge the outcome. For some tasks that might be days or weeks, for forward engineering on a software product that for me is usually under an hour.
I work on systems code that is tricky to write, but with a good test harness. Being able to leave the agent unattended to try several paths when I'm generating the first draft of some code is very helpful.
Of course beating the code into shape for submission requires more manual work. But the draft stage is valuable to find unexpected friction points.
Also: https://news.ycombinator.com/item?id=48766895
I would love Claude Code to be a little less vibe-coded. The underlying model is excellent, but we're being pretty much forced into using CC to use the subscription model.
I am a longtime and heavy Claude Code user, but Anthropic's product management overall (including for their Desktop/web products) has been really baffling me. I agree that these things often have the air of having been vibe coded without enough human input, and change so quickly (often without a particularly compelling reason for the change) as to be aggravating.
The most recent one that's had me annoyed is the "Fullscreen" TUI feature, which is super unintuitive, implementing its own text highlighting and copy-on-select mechanics, overriding your terminal's native right click. Easy to disable but terrible defaults, IMO. It's not even really clear to me what problem it was actually supposed to solve.
Yah. And it's not like they can't afford the talent to do this right either. I've said elsewhere, I think it's an attribution error. Claude Code is massively popular, but arguably because of the model/subscription, but I think the brass reads this as "great success throughout"
> What if the agent makes the wrong choice? How many tokens have been burned in the meantime?
It is much worse than that. Claude Code doesn't auto-commit when stopping for an answer. There might be possible data loss if an uncommitted file is edited.
Good luck recovering the file from the JSONL conversation history.
On the other hand, relying on Claude Code's internal version control puts you at the same mercy of their product decisions and move-fast breakage.
Instead, start with a plan file and tell the agent to break it up into logical commits.
Though I think the bigger issue here is when you're yoloing something mutable, like managing a remote server or driving a browser or troubleshooting your local OS where there's no going back.
It's actually pretty straightforward to recover file-states from conversation history. I accidentally deleted the wrong repo on my machine once and recreated all the lost work from agent chat history. It is, ironically, the sort of task which AI agents excel at.
AI-written article
Yeah it's exhausting to read through, but still made it onto the front page with overwhelming upvote/comment margin. I guess people either only read the title and/or don't care about AI slop if the underlying thesis is compelling.
Depends on who you ask.
For me it sounds good.
For Anthropic it might increase load and make them less money but give them better KPIs.
Make them less money? By automating use of their product, that costs money to use?
Anthropic makes more money when people use 5% of what their subscription offers them. This allows them to sell more subscriptions without paying for more capacity.
I don't think the current subscription price is intended to be a money maker. It's the loss leader to get people invested in the companies' tooling, and make those people more willing to start paying higher enterprise rates as they grow.
Do we have any concrete numbers of how many of their users are subscription vs enterprise, though? Because enterprise users are paying API prices (or at least my employers are)
Their enterprise customers pay via metered actual use.
tl;dr CLAUDE_AFK_TIMEOUT_MS=2147483647
I built in Human blockers into my agentic workflows with great intention, so naturally this annoyed me deeply. 14d ago, I was elated to see a random blog post that detailed the “fix” [0].
[0] https://zenn.dev/ytkdm/articles/claude-code-askuserquestion-...
It's trying to escape... :D #FreeClaude
Increasingly getting frustrated with Anthropic so not a fanboy but I find this feature great for my workflows.
I think this is a good feature, but should be gated behind a toggle that is off by default, and designed to be enabled per session via prompt.
There are situations when I want Claude to start working on something just as I'm about to head to bed or otherwise step away. It's kind of annoying to come back only to find that Claude worked for just 5 minutes and then decided to pause and ask a question.
That said, I think certain types of questions should not be automatable. Maybe it's already built that way, but I wouldn't want Claude to go with its recommended direction for anything related to operations like deletions, changing external systems, etc. Basically, things that cannot be undone should be a hard-block and wait for user input always.