I am glad articles like this are finally starting to get some momentum around what I call the LLM magic box industry. From caveman mode to RTK to semantic search and everything in between. Developers have become magicians that cast spells instead of engineers. It sucks at work especially with everyone so sure that their magic spell is the one for ultimate token savings.
My criteria are: if it’s not in a harness it’s probably not that good (the best ideas float up to Codex/Claude imo) and any GitHub advertising some percent of token savings is not to be trusted.
It’s hard to avoid the snake oil and I hope people start thinking critically on this stuff.
I burn a ton of tokens, but things actually have to prove their value. And the vast majority of things do not come close to doing so.
I have my own AI agent full of stuff. I blind A/B test everything, but I also don't think the results are all that useful as a signal to others.
Just because I Blind A/B test it 4 months ago, it's maybe not meaningful today.
Maybe the word choices I use dramatically impact things.
I do it, because I can prove the value, and see it with my own eyes. I don't even bother publishing the specific Blind A/B tests.
Also, I've seen other people try to Blind A/B test and get it very wrong. If your measurements aren't good, the test is meaningless.
I don't know. We're all working on these problems together. There's a lot of black magic (which is why I rely on hooks a lot). I'm sure I have tons of black magic, I have a large little AI Agent.
But what I know for certain, is it works for me. All it takes is for me to not use it, and I honestly don't know how everyone currently works with AI.
I will link it, but it is not an endorsement for what you do. Mostly only other software engineers use it. And it's so very specific to the things I have to do.
At best, maybe it sparks an idea for you to implement on your own.
I'll go further and note that some of the optimizations I've seen in rtk for things like `git status` have actually bubbled up into the model layer -- Codex is regularly making tool calls like `git status --short` instead of `git status`.
The idea itself is sound: If you can reduce the signal-to-noise ratio in the context window, then that's a good thing.
Whether or not RTK actually does this has not been established. I would be glad to see some proper benchmarks done on the actual difference this tool makes (not some meaningless "up to 90%" type of language).
I mean it kind of already is in harnesses. Codex and Claude Code both have subagent tools. You could probably get a similar token output cut just by asking Claude Code to run all commands with Haiku as a summarizing subagent.
The core of the problem is that there are a million tools that make AI better, and no ways to measure whether AI is working better.
Big companies with popular products have it. They do something between normal product analytics and chatbot evals to figure out if users are being successful in their sessions. That's the job.
But any given dev, with between 3 and 50 sessions a day? Like, I have no idea what makes the LLM better. It's all vibes.
My company has a whole stack here. Preferred harnesses, preferred models, skills, the shape of our code, everything. There's gotta be a way to measure whether this setup is working for us, at 1 / 1-million-th the scale of a Claude Code.
Tool use output represents a large amount of my output. I'll take 3.7M tokens saved on 3.9M tokens of input. Tokens saved are tokens saved.
> 3. Where Are the Accuracy Benchmarks?
As a user of RTK, it would be nice to see accuracy benchmarks. However, I've seen no evidence of the model missing anything critical as a result of the compression. As part of their design philosophy they are very strict about preserving correctness to the point that if a filter fails they fall back to raw output. For my most frequently used commands I've inspected the source, was happy with what I saw, they've earned my trust thus far.
> The day git, cargo, npm, or grep updates its terminal formatting by a few spaces or changes an error layout, RTK's regex and parsing filters will break. And returning to the silent failure trap, it won't throw an explicit error; it will fail quietly, feeding corrupted or partial text to your agent.
Again, any filter that fails simply falls back to the raw output. One of their core pillars is avoiding this exact scenario you described. RTK should never feed corrupted or partial text to an agent.
Your concerns are fair but I'd like to see your criticism backed up with evidence. Have you used RTK? Have you found evidence that they are failing to preserve correctness?
Fwiw, I just ran the steps to reproduce and got `Error: prettier produced no output` on rtk (0.42.2). Not saying this isn't valid for the users environment but I could not reproduce on linux.
Author of the text here. I will be honest with why I wrote it, the rtk ai looks very odd to me as software engineer, the number of stars, no mention of accuracy and how management is pushing that stuff to optimize costs. Now people are wrapping every possible command in rtk and trying to handle all major possible commands and decide which output you should get.
I tried it and it does not compress messages which was 90% of my context, so it only compresses a small part of my token usage. If you read it carefully you will realize that is exactly stated. If you look at /context you will probably see that tool calls are not where you are spending token on, so a proxy that compresses tool calls will not make much impact, whilst still being true that it compresses tool calls by 8x. Its just not that important for long coding sessions for me.
"native/built-in Read or cat tools, the data is not intercepted by RTK's shell hook"
I don't disagree with the article, but I also don't disagree with RTK. The output of these commands is not optimized for agents (or humans) for that matter.
I've been trying out RTK and it seems kinda alright. I doubt it's saving much, but the quality of the work feels similar.
But if it's making a dent in token usage (which I have not personally measured), then that's great.
I had to add some system prompt instructions to Pi to help it work (GPT 5.5 initially got confused when `git status` looked different than expected). The Claude Code extension appears to do a proper job of informing the agent about the unexpected shape of the output without any extra work on my part.
so how do you justify it's usage if it's not saving much and the work feels similiar. They have 664 issues open and some of them are quite funny, the tools are called and return success even though they aren't even installed.
My take is that handling so many versions and so many different tools shouldn't be the work of any single repo. The responsibility should be either on coding agent to compress or best case scenario people who are responsible for cli tool
I'm not justifying its usage, and I don't have to.
I've been trying it out for a couple days and it seems kinda OK or whatever. If that upsets you, then that's your problem.
I might dump it later on if it doesn't provide much if a benefit. I typically try out new things, then cull whatever doesn't work. This tool seems pretty neutral for now, at least.
no, it doesn't upset me. I am open for discussion, there might be things I miss and don't understand. I am just trying to get why it's been pushed so hard lately and if the benefits are really there. Sorry, if I sounded upset to you, but I am trying to be really civil and just genereally curious
Well, I'm sorry as well. I mistakenly assumed you were being confrontational.
There are a lot of people who have negative knee jerk reactions to any AI stuff, new workflows (I'll agree there is a lot of garbage being shilled in this space), etc., and I jumped the gun by lumping you into that group.
Nope, I am doing my master thesis on finetuning llms at 36, so I am into this stuff, but it’s been very weird lately. I’ve been self-taught dev and I definitely was missing computer science concepts so excited to fill the gaps, although the timing wasn’t perfect.
Good conversation! Great pushback against my arguments. That’s what I signed on with hacker news and missing that spirit recently
I feel like the state of the art is baked into the compaction logic, and I've had a lot of problems with compaction (absent other prompting) losing key bits of state.
https://github.com/toon-format/toon is another interesting one, and I feel like it takes on a much more achievable goal - reduce whitespace and verbosity of JSON, not overall context compression.
Personally, I find compaction to be unreliable, which forces me to rely heavily on session-specific planning documents and inter-agent handoff messages.
I don't agree with the conclusion at all. I can see the value of RTK - whether it is buggy or vibe coded is kind of secondary. That basically comes down to how severe and often the bugs are.
There's no gamification of savings here. Tool output can be meaty.
Is the author skeptical of the concept, or the implementation? Because only one of those is worth critiquing.
so burn more tokens to save more tokens, so that we can spend more on X token but save on Y tokens?
not the question is which X tokens and which Y tokens? and since the output is non-deterministic how do you validate this?
LLMs aren't random and that enforces something that people are too dumb to realize that random-ness could be normally distributed but LLMs have no reason to be normally distributed or follow any sort of curve of understanding.
They are non-deterministic but with bias so their output might be just be worse with T' transformation for the class of problems A is solving but work great for B. or vice versa.
You can't reproducibly test LLMs and that allows all sorts of benchmarks to exist which can make any model look good or bad as much as we want. Enlightening stuff.
Not much different from sociological or psychological sciences where with enough bias in data you can prove anything.
If you need a piece of information that is buried somewhere, or a high-level summary/distillation of a larger body of info, then subagents may be the right tool for the job.
If you need all the gathered context for later use (i.e. distilled context is insufficient), then subagents probably are not the right tool for the job.
First, I only say this because of what I learned as a phD inhuman memory, not as someone who authors agentic workflows or does AI.
How human cognition tends to work by simultaneously utilizing and combining/separating multiple frequency scales of information. A simple way of thinking about is this: We tend to encode and retrieve both the gist of what is happening, and the verbatim details of what happened. The gist can be thought of as low frequency information, almost like bullet points, that contain the big overview goal, keypoints). The verbatim traces, are the high resolution memory that contains all the details. The gist helps encoding and recall by providing encoding and retrieval context cues. There are also levels in between those two, but I was keeping it simple. During human development, verbatim memory capacity increases first, but then hits a wall/plateau. Further performance increases begin to depend on the ability to utilize and gain from gist-like representations that can guide encoding and retrieval of verbatim details within contexts.
You don't need to keep everything in the context window. My untested, perhaps naive hypothesis is that what is needed is that sub-agents dealing with verbatim tasks (actually writing code), their context window should be managed by an agent above that is tuned to information at a lower frequency, and it by another above it on even lower frequency information. Lowest frequency information context windows feel up slowly. High-frequency information fills up fast. Use the low frequency information to retrieve the needed high frequency information.
This is a really useful technique in my experience. The harnesses are starting to do it more on their own but if you encourage the use of more subagents, I find it's typically nothing but win.
thank you, author here. I will stay civil here and focus on the rtk as that was the goal of article.
So do you think rtk cli is ai slop? I had some suspicions looking at their repo and number of issues and their style. The prettier issue with running successfully while binary wasn't even installed was quite entertaining
I am glad articles like this are finally starting to get some momentum around what I call the LLM magic box industry. From caveman mode to RTK to semantic search and everything in between. Developers have become magicians that cast spells instead of engineers. It sucks at work especially with everyone so sure that their magic spell is the one for ultimate token savings.
My criteria are: if it’s not in a harness it’s probably not that good (the best ideas float up to Codex/Claude imo) and any GitHub advertising some percent of token savings is not to be trusted.
It’s hard to avoid the snake oil and I hope people start thinking critically on this stuff.
This is why I Blind A/B test everything.
I burn a ton of tokens, but things actually have to prove their value. And the vast majority of things do not come close to doing so.
I have my own AI agent full of stuff. I blind A/B test everything, but I also don't think the results are all that useful as a signal to others.
Just because I Blind A/B test it 4 months ago, it's maybe not meaningful today.
Maybe the word choices I use dramatically impact things.
I do it, because I can prove the value, and see it with my own eyes. I don't even bother publishing the specific Blind A/B tests.
Also, I've seen other people try to Blind A/B test and get it very wrong. If your measurements aren't good, the test is meaningless.
I don't know. We're all working on these problems together. There's a lot of black magic (which is why I rely on hooks a lot). I'm sure I have tons of black magic, I have a large little AI Agent.
But what I know for certain, is it works for me. All it takes is for me to not use it, and I honestly don't know how everyone currently works with AI.
I will link it, but it is not an endorsement for what you do. Mostly only other software engineers use it. And it's so very specific to the things I have to do.
At best, maybe it sparks an idea for you to implement on your own.
https://github.com/notque/vexjoy-agent
> if it’s not in a harness it’s probably not that good
Counterpoint: Everything under the sun is in some harness. Doom is in a harness — more than one, actually.
I'll go further and note that some of the optimizations I've seen in rtk for things like `git status` have actually bubbled up into the model layer -- Codex is regularly making tool calls like `git status --short` instead of `git status`.
The idea itself is sound: If you can reduce the signal-to-noise ratio in the context window, then that's a good thing.
Whether or not RTK actually does this has not been established. I would be glad to see some proper benchmarks done on the actual difference this tool makes (not some meaningless "up to 90%" type of language).
I was wondering if that impacts the accuracy, obviously the rtk output wasn't in the training dataset, but maybe it doesn't matter at the end
I have to say I made a similiar mistake with trusting semantic search is the next big thing. My opinion shifted, but it made sense for me for too long
I mean it kind of already is in harnesses. Codex and Claude Code both have subagent tools. You could probably get a similar token output cut just by asking Claude Code to run all commands with Haiku as a summarizing subagent.
There is a conflict of interest, though.
The core of the problem is that there are a million tools that make AI better, and no ways to measure whether AI is working better.
Big companies with popular products have it. They do something between normal product analytics and chatbot evals to figure out if users are being successful in their sessions. That's the job.
But any given dev, with between 3 and 50 sessions a day? Like, I have no idea what makes the LLM better. It's all vibes.
My company has a whole stack here. Preferred harnesses, preferred models, skills, the shape of our code, everything. There's gotta be a way to measure whether this setup is working for us, at 1 / 1-million-th the scale of a Claude Code.
> 1. Gamified Savings vs. Your Actual API Bill
Tool use output represents a large amount of my output. I'll take 3.7M tokens saved on 3.9M tokens of input. Tokens saved are tokens saved.
> 3. Where Are the Accuracy Benchmarks?
As a user of RTK, it would be nice to see accuracy benchmarks. However, I've seen no evidence of the model missing anything critical as a result of the compression. As part of their design philosophy they are very strict about preserving correctness to the point that if a filter fails they fall back to raw output. For my most frequently used commands I've inspected the source, was happy with what I saw, they've earned my trust thus far.
> The day git, cargo, npm, or grep updates its terminal formatting by a few spaces or changes an error layout, RTK's regex and parsing filters will break. And returning to the silent failure trap, it won't throw an explicit error; it will fail quietly, feeding corrupted or partial text to your agent.
Again, any filter that fails simply falls back to the raw output. One of their core pillars is avoiding this exact scenario you described. RTK should never feed corrupted or partial text to an agent.
Your concerns are fair but I'd like to see your criticism backed up with evidence. Have you used RTK? Have you found evidence that they are failing to preserve correctness?
I was looking through the issues as investigation. Some issues that caught my attention are looking quite bad https://github.com/rtk-ai/rtk/issues/2494 https://github.com/rtk-ai/rtk/issues/2462 https://github.com/rtk-ai/rtk/issues/2395
Fwiw, I just ran the steps to reproduce and got `Error: prettier produced no output` on rtk (0.42.2). Not saying this isn't valid for the users environment but I could not reproduce on linux.
appreciate the engineering effort and skin in the game. I might try on macos today as the author of issue.
Author of the text here. I will be honest with why I wrote it, the rtk ai looks very odd to me as software engineer, the number of stars, no mention of accuracy and how management is pushing that stuff to optimize costs. Now people are wrapping every possible command in rtk and trying to handle all major possible commands and decide which output you should get.
Why didn’t you offer any real world usage numbers to illustrate your point? I found this unhelpful.
I tried it and it does not compress messages which was 90% of my context, so it only compresses a small part of my token usage. If you read it carefully you will realize that is exactly stated. If you look at /context you will probably see that tool calls are not where you are spending token on, so a proxy that compresses tool calls will not make much impact, whilst still being true that it compresses tool calls by 8x. Its just not that important for long coding sessions for me.
"native/built-in Read or cat tools, the data is not intercepted by RTK's shell hook"
I don't disagree with the article, but I also don't disagree with RTK. The output of these commands is not optimized for agents (or humans) for that matter.
I've been trying out RTK and it seems kinda alright. I doubt it's saving much, but the quality of the work feels similar.
But if it's making a dent in token usage (which I have not personally measured), then that's great.
I had to add some system prompt instructions to Pi to help it work (GPT 5.5 initially got confused when `git status` looked different than expected). The Claude Code extension appears to do a proper job of informing the agent about the unexpected shape of the output without any extra work on my part.
so how do you justify it's usage if it's not saving much and the work feels similiar. They have 664 issues open and some of them are quite funny, the tools are called and return success even though they aren't even installed.
My take is that handling so many versions and so many different tools shouldn't be the work of any single repo. The responsibility should be either on coding agent to compress or best case scenario people who are responsible for cli tool
I'm not justifying its usage, and I don't have to.
I've been trying it out for a couple days and it seems kinda OK or whatever. If that upsets you, then that's your problem.
I might dump it later on if it doesn't provide much if a benefit. I typically try out new things, then cull whatever doesn't work. This tool seems pretty neutral for now, at least.
no, it doesn't upset me. I am open for discussion, there might be things I miss and don't understand. I am just trying to get why it's been pushed so hard lately and if the benefits are really there. Sorry, if I sounded upset to you, but I am trying to be really civil and just genereally curious
Well, I'm sorry as well. I mistakenly assumed you were being confrontational.
There are a lot of people who have negative knee jerk reactions to any AI stuff, new workflows (I'll agree there is a lot of garbage being shilled in this space), etc., and I jumped the gun by lumping you into that group.
Nope, I am doing my master thesis on finetuning llms at 36, so I am into this stuff, but it’s been very weird lately. I’ve been self-taught dev and I definitely was missing computer science concepts so excited to fill the gaps, although the timing wasn’t perfect.
Good conversation! Great pushback against my arguments. That’s what I signed on with hacker news and missing that spirit recently
I feel like the state of the art is baked into the compaction logic, and I've had a lot of problems with compaction (absent other prompting) losing key bits of state.
https://github.com/toon-format/toon is another interesting one, and I feel like it takes on a much more achievable goal - reduce whitespace and verbosity of JSON, not overall context compression.
Personally, I find compaction to be unreliable, which forces me to rely heavily on session-specific planning documents and inter-agent handoff messages.
I don't agree with the conclusion at all. I can see the value of RTK - whether it is buggy or vibe coded is kind of secondary. That basically comes down to how severe and often the bugs are.
There's no gamification of savings here. Tool output can be meaty.
Is the author skeptical of the concept, or the implementation? Because only one of those is worth critiquing.
Hey, author here, I am skeptical of implementation starting from Rust Token Killer and looking to monetize on Rust love by other developers.
Concept is fine to me and I believe we should optimize, but a repo that will handle all tools sounds like Sisyphus rolling a rock up the hill.
I feel like what is needed is not compression, but aggressive context management with subagents.
so burn more tokens to save more tokens, so that we can spend more on X token but save on Y tokens?
not the question is which X tokens and which Y tokens? and since the output is non-deterministic how do you validate this?
LLMs aren't random and that enforces something that people are too dumb to realize that random-ness could be normally distributed but LLMs have no reason to be normally distributed or follow any sort of curve of understanding.
They are non-deterministic but with bias so their output might be just be worse with T' transformation for the class of problems A is solving but work great for B. or vice versa.
You can't reproducibly test LLMs and that allows all sorts of benchmarks to exist which can make any model look good or bad as much as we want. Enlightening stuff.
Not much different from sociological or psychological sciences where with enough bias in data you can prove anything.
Use the right tool for the job.
If you need a piece of information that is buried somewhere, or a high-level summary/distillation of a larger body of info, then subagents may be the right tool for the job.
If you need all the gathered context for later use (i.e. distilled context is insufficient), then subagents probably are not the right tool for the job.
if your corebase requires a million tokens, then youre probably going to break more than you fix
If you are using a million tokens in a single context window, you are using the entire toolbox incorrectly.
I am the author the text.
What do you mean by aggresive context management with subagents? Would you add a lopp that would trim the context?
Both of those tasks seem even more difficult
First, I only say this because of what I learned as a phD inhuman memory, not as someone who authors agentic workflows or does AI.
How human cognition tends to work by simultaneously utilizing and combining/separating multiple frequency scales of information. A simple way of thinking about is this: We tend to encode and retrieve both the gist of what is happening, and the verbatim details of what happened. The gist can be thought of as low frequency information, almost like bullet points, that contain the big overview goal, keypoints). The verbatim traces, are the high resolution memory that contains all the details. The gist helps encoding and recall by providing encoding and retrieval context cues. There are also levels in between those two, but I was keeping it simple. During human development, verbatim memory capacity increases first, but then hits a wall/plateau. Further performance increases begin to depend on the ability to utilize and gain from gist-like representations that can guide encoding and retrieval of verbatim details within contexts.
You don't need to keep everything in the context window. My untested, perhaps naive hypothesis is that what is needed is that sub-agents dealing with verbatim tasks (actually writing code), their context window should be managed by an agent above that is tuned to information at a lower frequency, and it by another above it on even lower frequency information. Lowest frequency information context windows feel up slowly. High-frequency information fills up fast. Use the low frequency information to retrieve the needed high frequency information.
I believe they mean aggressive delegation to minimize context bloat in the coordinating agent.
This is a really useful technique in my experience. The harnesses are starting to do it more on their own but if you encourage the use of more subagents, I find it's typically nothing but win.
that would make more sense, trimming context with subagents sounds like an overkill
"Where Are the Accuracy Benchmarks?"
I wish the author would have provided one.
Am I the only one that thought RTK was Real-Time Kinematics used for precision with satellite navigation?
No. I clicked here for the same reason.
I might have picked better title, but they are literally called rtk https://github.com/rtk-ai/rtk
and it stands for Rust Token Killer
slop complaining about other slop
thank you, author here. I will stay civil here and focus on the rtk as that was the goal of article.
So do you think rtk cli is ai slop? I had some suspicions looking at their repo and number of issues and their style. The prettier issue with running successfully while binary wasn't even installed was quite entertaining
Did you use an LLM for the blog post? it reads like it in places.
I have raycast shortcut for fix grammar, it might done more damage than adding a, an, the or changing tenses.
A content-free 2nd "paragraph" like this turned me off immediately.
> But in the current dev tools gold rush, if something sounds too good to be true, it almost always is.
The people who are interested in RTK and in criticism of RTK aren't interested in pablum like this.
ok, this one is all mine. So that's even more hurtful as this is 100% me