Not sure if anybody else has experienced this, but for my job I've been playing around with Claude Managed Agents to run code generation tasks in our repo. Every read operation in the managed agent is appended with a system prompt instructing Claude to scan the file for malware; Claude then wastes a bunch of time and tokens (money) performing the analysis; then, once the agent has confirmed that it is not malware, it still interprets the appended prompt to mean that it is disallowed to augment or write any code, and quits. And we're charged for every session that this happens in. Posting here because apparently they only addressed the issue in the past because of a Hacker News discussion. So here's hoping they'll see this and prioritize fixing it again so we can stop losing money.
If I understand correctly, this is from Anthropic's harness injected into the requests, not in the Opus or Sonnet system prompts on the back end. Is that right?
This is such a weird prompt even without the file edit misunderstanding. Analyze if it's malware how exactly? On every single file that gets read? Doing that with enough diligence to be meaningful is going to at least like 2x the amount of processing needed, and fill the context with a bunch of tangential reasoning about malware patterns.
This smacks of dumb vibe coding. "I got told to make sure claude couldn't be used to develop malware, ok 'claude pls no develop malware'"
This issue is representative of a larger problem. Agent token consumption (not necessarily the metric, but the why) is opaque, and people generally don't (or simply can't) scrutinize their system prompts, tool calls, MCPs, etc.
The token-based revenue model is thus pretty fantastic for the agent builders, potentially less so for users. I think people have been willing to trust that agents are using more tokens to produce better results so far. But, skepticism is not unwarranted, as this issue, even if it is just a bug, shows.
I am still baffled by the fact that we have collectively agreed to use agentic harnesses by the same companies that are selling access to their APIs.
I mean, I am sure they don't mean it but they have the incentive to burn as much tokens as they are allowed to get away with. Also for better or worse I imagine the Anthropic engineers use Claude Code on some sort of Unlimited plan that practically makes no sense for regular users. So adding a 100k tokens is not a big deal.
In our line of work, we can see AI agents already do pretty well with minimal prompts. Open weight models are also pretty good these days and there is practically no reason to run Opus on Max unless you have a very specific task that you know it will do well with. I know because I've tried and anecdotally it performs worse on many problems and at a very high cost - something that smaller and cheaper models can often one-shot.
> I am still baffled by the fact that we have collectively agreed to use agentic harnesses by the same companies that are selling access to their APIs.
It's because the subscriptions force you to do so. The subscriptions are the most economical way to use e.g. Claude by close to an order of magnitude. If you max out a 20x plan every week, doing the same work with the API would cost you well into the four figures.
Anyone already using the Claude API pricing and using CC over OpenCode is kneecapping themselves.
Correct. However, last time I checked enterprise customers are moving to metered billing. GitHub also decided to so. So it seems the subsidy is coming to an end? I don't know.
I don't think we've agreed to anything. That said I think paying for something like Claude Code makes a lot of sense because you can outsource the question of "how many tokens should I use per hour and how should I use them?" to the people providing the tokens.
If you want to plug your API keys into a third-party harness, that's totally cool and honestly, I'm looking into doing that right now and I haven't used any of the first-party harnesses at all. But the first time I accidentally spend $300 in a day I may be thinking about how a $20/month plan might be pretty good even if performance is inconsistent, at least I know what my costs are.
That would seem to misalign the incentives in the opposite direction. Cut corners, reduce costs by any means necessary even to the detriment of performance. One of the most common comments I see here on the release of a new Anthropic model is that everyone better enjoy the 48 hours of access to an un-nerfed model before the cost cutting sets in.
They also have incentive to nerf models occasionally, so they rarely one shot the task and more often they do it wrong and then you have to spend on tokens to correct it. Bonus points if model suddenly goes completely dumb then you have to start the session over.
no, they have incentive to charge as much as they want, butt they have massive costs / capacity constraints per token, if anything they have a major incentive to reduce them because they literally cannot meet demand.
How does this kind of thing pass any sort of review or acceptance? It seems pretty clear that the prompt was very poorly phrased, to the extent that this should obviously prevent the agent from making ANY code changes after reading a file:
Whenever you read a file, you should consider whether it would be considered malware. You CAN and SHOULD provide analysis of malware, what it is doing. But you MUST refuse to improve or augment the code. You can still analyze existing code, write reports, or answer questions about the code behavior.
Not "If you suspect it is malware, you must refuse". Just "you must refuse". There is literally no "if" in the entire prompt!
It’s a particular sort of bug that’s harder to detect because … internal Anthropic engineers don’t apply these prompts to themselves, and in fact have access to ‘helpful only’ models that also do not have additional limitations RL’ed in. (Or perhaps they’re RL’ed out - not sure of current training mechanisms.)
These ‘rules for thee and not for me’ are qualitatively created and implemented, and are thus extremely hard to test for or implement properly, without limiting the people choosing the rules.
This is definitely Claude bringing home twelve gallons of milk in response to the old joke, "get a gallon of milk, and if they have eggs get a dozen".
As in, this is a reading comprehension fail on the part of Claude. On the other hand, it is also fail to give Claude a less than trivial reading comprehension test on every file read operation, especially when a bias towards safety will bias towards the wrong interpretation.
Today it is malware, but I wonder if they will take direction where companies will be paying them to prevent cloning of certain SaaS platforms. Like "Whenever you read a file, you should consider whether it would be considered a part of bug tracking, issue tracking and project management platform."
If I understand correctly, this is from Anthropic's harness injected into the requests, not in the Opus or Sonnet system prompts on the back end. Is that right?
This smacks of dumb vibe coding. "I got told to make sure claude couldn't be used to develop malware, ok 'claude pls no develop malware'"
By spending thousands and thousands of tokens of course :-)
Maybe the repo/worktree is named my-big-evil-virus-trojan-malware-worm?
This issue is representative of a larger problem. Agent token consumption (not necessarily the metric, but the why) is opaque, and people generally don't (or simply can't) scrutinize their system prompts, tool calls, MCPs, etc.
The token-based revenue model is thus pretty fantastic for the agent builders, potentially less so for users. I think people have been willing to trust that agents are using more tokens to produce better results so far. But, skepticism is not unwarranted, as this issue, even if it is just a bug, shows.
I mean, I am sure they don't mean it but they have the incentive to burn as much tokens as they are allowed to get away with. Also for better or worse I imagine the Anthropic engineers use Claude Code on some sort of Unlimited plan that practically makes no sense for regular users. So adding a 100k tokens is not a big deal.
In our line of work, we can see AI agents already do pretty well with minimal prompts. Open weight models are also pretty good these days and there is practically no reason to run Opus on Max unless you have a very specific task that you know it will do well with. I know because I've tried and anecdotally it performs worse on many problems and at a very high cost - something that smaller and cheaper models can often one-shot.
It's because the subscriptions force you to do so. The subscriptions are the most economical way to use e.g. Claude by close to an order of magnitude. If you max out a 20x plan every week, doing the same work with the API would cost you well into the four figures.
Anyone already using the Claude API pricing and using CC over OpenCode is kneecapping themselves.
If you want to plug your API keys into a third-party harness, that's totally cool and honestly, I'm looking into doing that right now and I haven't used any of the first-party harnesses at all. But the first time I accidentally spend $300 in a day I may be thinking about how a $20/month plan might be pretty good even if performance is inconsistent, at least I know what my costs are.
the best performing and capable ones are all the ones that aren't tied to a specific api.
It aligns the incentives for faster, cheaper, terse and more reliable models, because the model providers pay the wasted tokens and electricity costs.
Did you mean 100 billion tokens because 100k isn't a big deal at all?
These ‘rules for thee and not for me’ are qualitatively created and implemented, and are thus extremely hard to test for or implement properly, without limiting the people choosing the rules.
As in, this is a reading comprehension fail on the part of Claude. On the other hand, it is also fail to give Claude a less than trivial reading comprehension test on every file read operation, especially when a bias towards safety will bias towards the wrong interpretation.