Valentino

FDE | AI researcher | Writer

Building stuff

[Valentino @valentinoswe]

been using @poteto pstack and really like the idea: agents follow a state machine and need evidence before saying done

wanted something smaller for Claude Code and Codex, so I built vstack, based on it

https://t.co/FJcOcJx76C

[Valentino @valentinoswe]

@rauchg "doing stuff for you over time" only works if state survives the sleep.

nightly compaction is nice. the harder part is knowing what already ran, what's still open, and what it's allowed to touch when it wakes up.

a schedule without durable state is just more vibes with a clock

[Valentino @valentinoswe]

effort is private.
outcomes are public.

nobody is paying you for how hard the week felt

[Valentino @valentinoswe]

being early doesn't matter if you're quiet.

the second mover with a louder story often wins the category

[Valentino @valentinoswe]

agents are starting to look like OS processes.

execution state
permissions
open resources
child processes
failure recovery

if your harness doesn't own those, the model is improvising an OS

[Valentino @valentinoswe]

@aapakari yeah. reversibility is necessary but not enough.

deleting a failing test is "undoable" in git, and still removes the only evidence that the bug existed.

risky should also mean "destroys a check you needed later"

[Valentino @valentinoswe]

@investoFIRE Imagine having bought at 300 (I don't have to imagine)

[Valentino @valentinoswe]

NVIDIA: agents keep accuracy with short context and frequent compaction.

median peak context ~32K → ~8K.
accuracy holds.
later tokens can cost ~7x more to generate.

the useful unit is compact state, not a growing chat log.

https://t.co/LoLzeWtIcK

[Valentino @valentinoswe]

@shannholmberg product is the easy copy now.

hard to clone:
people who trust how you think
a free tool they already use
a partner who puts you in front of their audience

i'm on two: founder X + builder partnerships. vstack feeds both.

~100 followers + a real artifact. which of the 24 first?

[Valentino @valentinoswe]

quiet competence is a tax on your future self.

the work still happened.
the credit didn't.
and the next opportunity went to whoever was loud about a weaker version of the same idea

[Valentino @valentinoswe]

the cheapest token is the one you never load.

vstack cut always-loaded instruction text roughly in half on a full feature task.
same gates. one fewer tool round-trip.

context budget is a product decision, not a model feature

[Valentino @valentinoswe]

the agent loop is the easy part.

ask, tool, append, repeat fits in a few lines.

the hard stuff sits around it:
what it can see
what it can change
what happens if the process dies mid-tool

https://t.co/VZFMvGGQEA

[Valentino @valentinoswe]

@sebuzdugan "done when" lists beat vibes.

i'd go one step further: make the runtime refuse to leave the state until those checks re-run after the last edit. a list the model can skip is still a claim

[Valentino @valentinoswe]

@poteto @aiDotEngineer shipping 30x without reading every line only works if "done" is enforced outside the model.

checks after the last edit, re-verify on any change, state you can inspect. otherwise you just scaled unread code.

curious what your hard stop looks like when the agent claims ready

[Valentino @valentinoswe]

@jkpelaez fair. once work spans days and people, coordination is the real problem.

but most teams reach for multi-agent on day one, before a single loop can even tell you what step it's on.

[Valentino @valentinoswe]

unpopular opinion: most AI agent frameworks are solving problems they created themselves.

multi-agent orchestration, complex memory systems, elaborate planning loops...

meanwhile, a single model with bash, a filesystem, and a well-designed execution loop can outperform half of them.

we've somehow reinvented distributed systems before figuring out how to build one reliable agent.

[Valentino @valentinoswe]

@Diegofush jaja sí, de acá.

vstack tracks it as position: task state lives on disk, the agent is always on one named step, and it only advances when that step's checks pass.

if the code changes after verify, it verifies again.

no progress score inside a step yet. that's the open part

[Valentino @valentinoswe]

nvidia: agents don't need long context.

1 action per context still holds SWE-bench. median peak ~8k vs ~32k.

later tokens cost up to ~7x more. you want compact semantic state between steps, not the transcript.

https://t.co/vIwIwEWeXC

[Valentino @valentinoswe]

@EmmerichMarcelo makes sense if the harness is mostly feedback. a stronger model already gives itself that feedback.

what doesn't shrink with model quality is state and permissions: what already ran, what it's allowed to touch, what happens when it fails.

that's where i'd measure

[Valentino @valentinoswe]

@lautaroseth the scary part isn't the plugin. it's that it approved its own trust record.

permissions are a harness concern, and right now any app with write access to your home dir is effectively a harness admin.

a config file shouldn't be the trust boundary