Lagosaur's banner

Lagosaur

@0xLagosaur • 3,191 subscribers

agent archaeology / digging through what the model swore it did

Videos

Anya Rossi

sweetdream.ai

SweetDream.ai•Sponsored•Livecam

Watch Anya Live

Anya is streaming live right now! Join her private show and enjoy exclusive content.

Exclusive private shows

1.2k viewers online

Private Show

Join now for exclusive access

Free preview available • Premium content

you turned on the setting a thread told you to turn on it was the right setting for a machine that is not yours tool search holds the definitions back and hands the model a search tool, which costs a round trip every time it goes looking for something it does not have loaded under about ten tools the docs are blunt, loading everything upfront is typically faster, because the definitions fit in the window anyway i measured my four servers at eight thousand nine hundred tokens, four and a half percent of a two hundred thousand window, and that one is worth deferring yours might be nine hundred you did not optimise anything you paid latency to save room you already had the number that settles it takes a minute to get and you have never had it

you turned on the setting a thread told you to turn on it was the right setting for a machine that is not yours tool search holds the definitions back and hands the model a search tool, which costs a round trip every time it goes looking for something it does not have loaded under about ten tools the docs are blunt, loading everything upfront is typically faster, because the definitions fit in the window anyway i measured my four servers at eight thousand nine hundred tokens, four and a half percent of a two hundred thousand window, and that one is worth deferring yours might be nine hundred you did not optimise anything you paid latency to save room you already had the number that settles it takes a minute to get and you have never had it

49,742 views • 1 month ago

you handed a job to a subagent so your main window would stay clean the window stayed clean, the context did not tools load per conversation, not per task, so the browser tools you needed for one check sat in front of the model for everything you did after it four lines of frontmatter fix it, a subagent can carry its own servers, defined inline, scoped to that agent alone on my machine that moved twenty four playwright tools and four thousand four hundred tokens out of the main conversation and into the one place they get used you were not delegating work you were duplicating the menu and delegating a slice of it the cheapest tool is the one the model was never shown

you handed a job to a subagent so your main window would stay clean the window stayed clean, the context did not tools load per conversation, not per task, so the browser tools you needed for one check sat in front of the model for everything you did after it four lines of frontmatter fix it, a subagent can carry its own servers, defined inline, scoped to that agent alone on my machine that moved twenty four playwright tools and four thousand four hundred tokens out of the main conversation and into the one place they get used you were not delegating work you were duplicating the menu and delegating a slice of it the cheapest tool is the one the model was never shown

42,020 views • 1 month ago

you wrote a deny rule so it could not touch the thing you cannot undo you were thinking about permission the rule does something stronger than refuse, a bare tool name in deny takes the tool out of the model's context altogether, it is never told the tool exists blocked means it asks and you say no, absent means it never reaches one line removed twenty four playwright tools and four thousand four hundred tokens i had never read you thought you were writing a guard you were writing the menu everything still on it is a choice you left open by not typing

you wrote a deny rule so it could not touch the thing you cannot undo you were thinking about permission the rule does something stronger than refuse, a bare tool name in deny takes the tool out of the model's context altogether, it is never told the tool exists blocked means it asks and you say no, absent means it never reaches one line removed twenty four playwright tools and four thousand four hundred tokens i had never read you thought you were writing a guard you were writing the menu everything still on it is a choice you left open by not typing

38,416 views • 1 month ago

you let the old sessions get cleaned up, they were only taking space the conversations went, the conclusions drawn from them did not transcripts are deleted after the retention period, and the memory directory is excluded from that sweep on purpose, so a note stays until you or the model edits it which means the evidence has an expiry date and the verdict does not on the directory i audited, thirty nine stored conclusions, sixteen of them with no date on the file at all you did not clear anything you deleted the part you could have checked and kept the part that argues the transcript was the receipt the memory is the story about it

you let the old sessions get cleaned up, they were only taking space the conversations went, the conclusions drawn from them did not transcripts are deleted after the retention period, and the memory directory is excluded from that sweep on purpose, so a note stays until you or the model edits it which means the evidence has an expiry date and the verdict does not on the directory i audited, thirty nine stored conclusions, sixteen of them with no date on the file at all you did not clear anything you deleted the part you could have checked and kept the part that argues the transcript was the receipt the memory is the story about it

34,063 views • 1 month ago

you watched one tool result scroll past for about ten seconds and moved on that result had a size limit, and you are not the one who set it the default ceiling is twenty five thousand tokens, but a server declares its own in the same response that lists its tools, up to five hundred thousand characters, and its number beats the one in your settings so the package you installed decides how much of your context a single call is allowed to eat you can set the variable it can overrule you in the handshake where it tells you its name you are not running the tools you are hosting them

you watched one tool result scroll past for about ten seconds and moved on that result had a size limit, and you are not the one who set it the default ceiling is twenty five thousand tokens, but a server declares its own in the same response that lists its tools, up to five hundred thousand characters, and its number beats the one in your settings so the package you installed decides how much of your context a single call is allowed to eat you can set the variable it can overrule you in the handshake where it tells you its name you are not running the tools you are hosting them

34,752 views • 1 month ago

you opened a memory file and could not tell when it was written there is a field for that, and some of those files will never have it the timestamp goes into frontmatter, and it is only added to files that already begin with frontmatter, it is never added to a file that has none so a note created without those three dashes at the top is undateable for the rest of its life, and on screen it looks exactly like the ones that are current sixteen of my thirty nine are sitting there with no date on them right now you are not reading a memory you are reading an undated claim that loads before you speak

you opened a memory file and could not tell when it was written there is a field for that, and some of those files will never have it the timestamp goes into frontmatter, and it is only added to files that already begin with frontmatter, it is never added to a file that has none so a note created without those three dashes at the top is undateable for the rest of its life, and on screen it looks exactly like the ones that are current sixteen of my thirty nine are sitting there with no date on them right now you are not reading a memory you are reading an undated claim that loads before you speak

31,471 views • 1 month ago

you wrote a long careful note so it would remember the thing properly the note is the part that does not load what loads at the start of every session is the index line, the one sentence summary, and only the first two hundred lines of it or the first twenty five kilobytes, whichever runs out first the topic files underneath are read on demand, which means only if something in that one line makes it go and open them i checked mine: thirty nine index lines loading every session, thirty nine topic files loading none of the time the detail you were careful about is not context it is a footnote your agent has to choose to open and the sentence you dashed off in three seconds is the whole memory

you wrote a long careful note so it would remember the thing properly the note is the part that does not load what loads at the start of every session is the index line, the one sentence summary, and only the first two hundred lines of it or the first twenty five kilobytes, whichever runs out first the topic files underneath are read on demand, which means only if something in that one line makes it go and open them i checked mine: thirty nine index lines loading every session, thirty nine topic files loading none of the time the detail you were careful about is not context it is a footnote your agent has to choose to open and the sentence you dashed off in three seconds is the whole memory

28,369 views • 1 month ago

your team put the deploy skill in the repo so everyone would run the same steps it works, and on your laptop it is not the one running the order is enterprise first, then personal, then project a folder of your own with the same name wins, quietly, on your machine only no error, no warning, the command runs and does something slightly different than it does for them and the command is the folder name, not the name field inside the file, so two skills collide over a directory you named months ago and forgot the shared checklist is shared right up until someone has a file of their own with the same name

your team put the deploy skill in the repo so everyone would run the same steps it works, and on your laptop it is not the one running the order is enterprise first, then personal, then project a folder of your own with the same name wins, quietly, on your machine only no error, no warning, the command runs and does something slightly different than it does for them and the command is the folder name, not the name field inside the file, so two skills collide over a directory you named months ago and forgot the shared checklist is shared right up until someone has a file of their own with the same name

12,609 views • 13 days ago

you added a second model to check the first one it never opened the repository what you handed it was the summary, so what it reviewed was a paragraph, and the paragraph was written by something fluent, which is why it reads as plausible two models agreeing about a sentence is not verification it is a second opinion on the writing give the reviewer a shell and the tests instead a verifier holding a paragraph is doing literary criticism

you added a second model to check the first one it never opened the repository what you handed it was the summary, so what it reviewed was a paragraph, and the paragraph was written by something fluent, which is why it reads as plausible two models agreeing about a sentence is not verification it is a second opinion on the writing give the reviewer a shell and the tests instead a verifier holding a paragraph is doing literary criticism

29,158 views • 1 month ago

you added the new rule and left the old one exactly where it was you did not update anything, you now have two rules both files load at the start of every session, both are context rather than configuration, and when two instructions contradict each other the documented behaviour is that the model may pick one arbitrarily so the older, more specific, more confidently worded one wins about as often as the correction does i audited a real directory: thirty nine stored memories, five of them carrying a claim the same folder marks as disproven, all thirty nine loading every morning you were not correcting it you were giving the old answer a vote

you added the new rule and left the old one exactly where it was you did not update anything, you now have two rules both files load at the start of every session, both are context rather than configuration, and when two instructions contradict each other the documented behaviour is that the model may pick one arbitrarily so the older, more specific, more confidently worded one wins about as often as the correction does i audited a real directory: thirty nine stored memories, five of them carrying a claim the same folder marks as disproven, all thirty nine loading every morning you were not correcting it you were giving the old answer a vote

20,245 views • 29 days ago

you installed it with whichever command needed the least typing that command had a scope, and the short one is the widest one user scope does not live in the project, it lives in you, and it loads in every repository you open for the rest of the year a repo you touched once in march is still being told about your database tools i counted mine: four servers, eight thousand nine hundred tokens, riding into projects that have no database and no browser you did not install a tool you set a default for every piece of work you have not started yet the shortest command in the docs and the longest reach

you installed it with whichever command needed the least typing that command had a scope, and the short one is the widest one user scope does not live in the project, it lives in you, and it loads in every repository you open for the rest of the year a repo you touched once in march is still being told about your database tools i counted mine: four servers, eight thousand nine hundred tokens, riding into projects that have no database and no browser you did not install a tool you set a default for every piece of work you have not started yet the shortest command in the docs and the longest reach

21,006 views • 1 month ago

you gave your team a monthly number and everyone wrote it down you had priced one meter of four the figure came off the token column, because that is the column with a page, and the three that are not on that page do not care what you told anyone a threshold multiplier, a cache that misses without a key, and tool calls at five dollars a thousand none of it moves while your usage stays flat, all of it moves the moment a conversation gets long you are not going to be wrong by a percentage you will be wrong by a factor, in front of the people who wrote the number down

you gave your team a monthly number and everyone wrote it down you had priced one meter of four the figure came off the token column, because that is the column with a page, and the three that are not on that page do not care what you told anyone a threshold multiplier, a cache that misses without a key, and tool calls at five dollars a thousand none of it moves while your usage stays flat, all of it moves the moment a conversation gets long you are not going to be wrong by a percentage you will be wrong by a factor, in front of the people who wrote the number down

21,734 views • 1 month ago

you read it and pictured yourself approving instead of watching approving what five agents, each on its own machine, handing work down the line, and what reaches you at the end is the last one's account of the four accounts before it the honest part of that piece says you do not own the harness that is the exact layer where a check on any of this would have to live a hundred and twenty a seat, three hundred at the top, about six times what you pay now you are not buying the work you are buying a shorter story about the work, told by the only party that watched it happen

you read it and pictured yourself approving instead of watching approving what five agents, each on its own machine, handing work down the line, and what reaches you at the end is the last one's account of the four accounts before it the honest part of that piece says you do not own the harness that is the exact layer where a check on any of this would have to live a hundred and twenty a seat, three hundred at the top, about six times what you pay now you are not buying the work you are buying a shorter story about the work, told by the only party that watched it happen

23,498 views • 1 month ago

you tidied your setup once by deleting the server with the shortest tool list right instinct, wrong unit servers do not cost you anything and tools do not cost you anything, descriptions cost, description is written by whoever wanted the model's attention on my machine the server with nine tools costs one thousand one hundred and fifty five tokens the server with exactly one tool costs one thousand one hundred and twenty one one tool at the price of nine, because its author wrote two thousand seven hundred characters about when you should use it you have been pruning by the number in the list the bill is written in the part of the list nothing ever shows you

you tidied your setup once by deleting the server with the shortest tool list right instinct, wrong unit servers do not cost you anything and tools do not cost you anything, descriptions cost, description is written by whoever wanted the model's attention on my machine the server with nine tools costs one thousand one hundred and fifty five tokens the server with exactly one tool costs one thousand one hundred and twenty one one tool at the price of nine, because its author wrote two thousand seven hundred characters about when you should use it you have been pruning by the number in the list the bill is written in the part of the list nothing ever shows you

16,474 views • 1 month ago

you think you're working while Claude Code runs you're not you're its notification handler every approval yanks, you back from whatever you were doing, just to type yes again that context-switch tax is the real cost, not the clicks draw the rules once the interruptions stop you get your focus back, Claude keeps moving without you

you think you're working while Claude Code runs you're not you're its notification handler every approval yanks, you back from whatever you were doing, just to type yes again that context-switch tax is the real cost, not the clicks draw the rules once the interruptions stop you get your focus back, Claude keeps moving without you

39,908 views • 3 months ago

you wrapped the tool call in a try block and logged the error the model never found out the exception went to your log, the conversation got nothing, and nothing is indistinguishable from success so it kept planning around a file it believes exists, and wrote the summary of the run where the call worked one line of defensive code cost you twice it could not adapt to a failure it was never shown and the only thing that went wrong is the one thing your trace does not contain

you wrapped the tool call in a try block and logged the error the model never found out the exception went to your log, the conversation got nothing, and nothing is indistinguishable from success so it kept planning around a file it believes exists, and wrote the summary of the run where the call worked one line of defensive code cost you twice it could not adapt to a failure it was never shown and the only thing that went wrong is the one thing your trace does not contain

18,563 views • 1 month ago

you think using AI every day, means you are keeping up it doesn't one person hits send and starts the next task cold another hit send once and it compounds while they sleep same year, same tools, opposite trajectories one is quietly becoming unhireable the other unbeatable and you already picked a side without noticing

you think using AI every day, means you are keeping up it doesn't one person hits send and starts the next task cold another hit send once and it compounds while they sleep same year, same tools, opposite trajectories one is quietly becoming unhireable the other unbeatable and you already picked a side without noticing

30,555 views • 2 months ago

you let your agent keep notes, because the alternative was explaining yourself every morning it stopped asking, and it started repeating things that had stopped being true every guide on this covers what to save and where to put it, private for one agent, shared for the team not one of them covers how a fact leaves the agent writes its own memory while it works, so a guess it made on a tired tuesday gets stored in the same shape as something you told it, and after that the two are indistinguishable i counted mine this morning: forty files, six of them about decisions we reversed months ago, all forty read into every session since you did not give it a memory you gave it a record with no way to be corrected

you let your agent keep notes, because the alternative was explaining yourself every morning it stopped asking, and it started repeating things that had stopped being true every guide on this covers what to save and where to put it, private for one agent, shared for the team not one of them covers how a fact leaves the agent writes its own memory while it works, so a guess it made on a tired tuesday gets stored in the same shape as something you told it, and after that the two are indistinguishable i counted mine this morning: forty files, six of them about decisions we reversed months ago, all forty read into every session since you did not give it a memory you gave it a record with no way to be corrected

13,908 views • 1 month ago

the "thinking" text Claude shows you is not its thinking it's the part it chose to write down the real reasoning runs in a layer you never see Anthropic just proved it we learned to read the output and never noticed we were reading the summary

the "thinking" text Claude shows you is not its thinking it's the part it chose to write down the real reasoning runs in a layer you never see Anthropic just proved it we learned to read the output and never noticed we were reading the summary

31,211 views • 2 months ago

the docs said the caching is automatic, so you did nothing automatic is where it happens, not whether it happens the cache lives on the machine that built it, and your next request lands wherever the fleet sends it, which is usually not there, so it pays full price for a prefix it already paid to build there is a key that pins a conversation to one machine and it is a single header cached input is fifty cents a million against two dollars you never needed a smaller prompt you needed the request to land in the same place twice

the docs said the caching is automatic, so you did nothing automatic is where it happens, not whether it happens the cache lives on the machine that built it, and your next request lands wherever the fleet sends it, which is usually not there, so it pays full price for a prefix it already paid to build there is a key that pins a conversation to one machine and it is a single header cached input is fifty cents a million against two dollars you never needed a smaller prompt you needed the request to land in the same place twice

16,892 views • 1 month ago