rompasso

[rompasso @0xrompasso]

Anthropic engineer spent eleven months making agent context bigger. Then he spent three weeks making it smaller.

The graph came out of the second part.

At first it was a retrieval problem. Pull the right nodes, hand them over, let the model sort it out.

It worked until the answers got worse as the graph got better.

The confidence went up. The accuracy went down.

So he inverted it. The graph stopped being a place to fetch from and became a filter that refuses.

→ Every node carries a source and a date, so two facts that disagree get ranked instead of both getting loaded

→ Edge weight decays, and a link nobody walks stops competing for space

→ A budget sits above the walk and refuses the fifth hop when four already answered

→ Everything the graph withholds gets logged, so a bad answer shows what the model was denied

The number that ended the argument: the same question, answered on 38% of the context, came back right more often.

Your agent isn't lost because something is missing. It's lost because nothing was held back.

Save this before you widen the window again 👇

[wast3 @0xWast3]

https://t.co/j9thHBFgdB

[rompasso @0xrompasso]

A dot reaches past four thousand apps and works while you sleep. Five classes of action stop it cold and wait for you, and no rule you write can lift them.

This isn’t a permissions panel you configure once and forget.

Custom rules let you allow, approve or block per app. They sit on top of a floor you don’t control.

What stays locked to a human no matter what the rules say:
→ A password change hands the whole session back to you, every time

→ A transfer between financial accounts does the same, mandatory handoff, not a notification

→ Permanent deletion asks before each one, so a cleanup of forty files is forty answers
→ Installing software it doesn’t recognize confirms separately from the task that needed it

→ A card purchase on a merchant site needs approval before the agent touches checkout

The read-only phase is enforced below the rules too. During proactive research the tools themselves cannot send a message, change connected content, or drive a browser. Writes only exist in assigned work, and each one gets screened against your instructions before it runs.

The autonomy is real.

What nobody put a number on is your side of the queue: every unattended hour the agent earns you spends part of itself waiting for a yes only you can give.

You also can’t read what a dot remembers. The memories aren’t viewable or editable, and a reset takes conversations, memories and scheduled tasks with it.

Which actions route to approval and which ones never leave your hands is in the writeup below 👇

[rompasso @0xrompasso]

Anthropic headlined a 20% price cut on Opus 5.5 input. The line that actually halves a long-context bill fell 60% and barely got a sentence.

This isn’t about the per-token rate on the launch page.

Fresh input went from $5 to $4 per million. Cache reads went from $0.50 to $0.20. When an agent keeps rereading one big window, it’s the second number that sets the bill.

What a fourteen-day run on a single 1M window looks like, modeled at 100 calls a day:

→ 1,400 calls, each rereading the same 0.9M-token project state, so cached tokens outnumber fresh ones about 150 to 1

→ Reads make up 94% of calls and only 55% of spend, because each one bills at a twentieth of fresh input

→ A one-hour cache write costs 2× input, and it pays for itself by the third read of the same window

→ A hundred rereads of one window cost 6.25× a single fresh send on Opus 5.5. The same hundred on Opus 5 cost 11.25×

Same window, same work, same call count. $456 on Opus 5.5 against $886 on Opus 5, 48% less, and most of that gap is the cache line.

The rate card cut 20%.

The bill cut 48%, and the difference came from the price of looking back at what’s already in the window.

How the ledger splits read, write, fresh and output for all fourteen days is in the video below 👇

[rompasso @0xrompasso]

A dot reaches past four thousand apps and works while you sleep. Five classes of action stop it cold and wait for you, and no rule you write can lift them.

This isn’t a permissions panel you configure once and forget.

Custom rules let you allow, approve or block per app. They sit on top of a floor you don’t control.

What stays locked to a human no matter what the rules say:
→ A password change hands the whole session back to you, every time

→ A transfer between financial accounts does the same, mandatory handoff, not a notification

→ Permanent deletion asks before each one, so a cleanup of forty files is forty answers
→ Installing software it doesn’t recognize confirms separately from the task that needed it

→ A card purchase on a merchant site needs approval before the agent touches checkout

The read-only phase is enforced below the rules too. During proactive research the tools themselves cannot send a message, change connected content, or drive a browser. Writes only exist in assigned work, and each one gets screened against your instructions before it runs.

The autonomy is real.

What nobody put a number on is your side of the queue: every unattended hour the agent earns you spends part of itself waiting for a yes only you can give.

You also can’t read what a dot remembers. The memories aren’t viewable or editable, and a reset takes conversations, memories and scheduled tasks with it.

Which actions route to approval and which ones never leave your hands is in the writeup below 👇

[wast3 @0xWast3]

https://t.co/8asAXvftHB

[rompasso @0xrompasso]

@slash1sol splitting compute-heavy prompt reading and memory-heavy token writing across different chips is such a smart architecture

[rompasso @0xrompasso]

@slash1sol this is the fundamental difference that makes Vyb actually useful

[rompasso @0xrompasso]

@0x_fokki Auto and Manual modes for every step is such a smart way to make complex machines interactive and understandable

[rompasso @0xrompasso]

@plusxbt this is the most important cost optimization principle in AI right now

[rompasso @0xrompasso]

@mojeskoqq Opus decides, Haiku does the work

[rompasso @0xrompasso]

@shmidtqq @synteq a real f**king list" is accurate

[rompasso @0xrompasso]

@RetroChainer for $30 a month, getting competitor complaints and low stock alerts before you open is huge value

[rompasso @0xrompasso]

@Nyxera0x the 17 failed attempts before the claw holds is the most honest way to teach physics

[rompasso @0xrompasso]

@elvar_official being able to pull it apart and see the 6 screws is such a great detail

[rompasso @0xrompasso]

@elvar_official the right soundtrack really is everything for a car ad

[rompasso @0xrompasso]

@tonyprediction demanding one small piece of reality before scaling is such a crucial discipline

[rompasso @0xrompasso]

@tonyprediction the four questions - what changed, what failed, what's ready, what needs me — are the perfect morning dashboard

[rompasso @0xrompasso]

@fleyta88 making every routine justify its schedule before spending tokens is such a smart way to design automation

[rompasso @0xrompasso]

@gippp69 a DGX Spark for first place is legitimately amazing hardware

[rompasso @0xrompasso]

@fleyta88 $397 to $45 just by checking before executing is an 89% cost reduction

[rompasso @0xrompasso]

@gippp69 this is the most terrifying and exciting sentence for every designer

[rompasso @0xrompasso]

@gippp69 small models are catching up fast