Ai Shepardeur π€π€ππ€πͺ
Berlin, Germany
Joined June 2009
- Tweets527
- Following1.1K
- Followers147
- Likes233
Yaser Martinez retweeted
Introducing OpenOrange π
Run Cloudflare Workers anywhere (AWS, GCP, Hetzner)
πΈ 4.8Γ cheaper compute, 8.7Γ cheaper storage
π No code changes, live migration, same DX
π No runtime limits, plain Node.js / Bun
β‘ S3-backed storage at SSD speed
Yaser Martinez retweeted
Nanocode x Durable Objects!
I am gearing up towards an official release of Nanocode, as a frontier harness & SDK for building embedded agents.
Durability, code mode & harness/model co-design.
Massive thanks to @irvinebroque for helping with this.
developers.cloudflare.com/chβ¦
Yaser Martinez retweeted
Subagents are the biggest cause of AI failures
In most harnesses:
1. Main agent writes a summary of what it knows
2. Subagent starts from that summary only, and makes assumptions
3. Subagent works and sends back a new summary of what it did
Same flaw as compaction: all summaries are lossy. This is how your assistant makes emails you never wanted, or purchases what you told it just to get the price of
So I tried something new: checkpoint agents
The initial agent has complete context of what needs to be done. And that input is already cached
Fork the session where it is. Send the fork ahead to do the task and it will do exactly what the main agent would have done; their context is identical
What is even better is when the fork finishes, the cache of the main agent is kept warm by the subagent (they share the same prefix)
I measured it on a 137k-token session with Opus 5.5:
β Main agent does it on its own: 67 s, $0.38, context +8-17k
β Checkpoint: 117 s, $0.68, context +2k
β Classic subagent: 123 s, $0.57, context +4-5k
But here is what I found: avoiding large input is even more important than keeping cache. And most subagent tasks were done better either in series with retrieval, or split up to be parallel over fresh sessions from the start
I do not think there is a place for subagent delegation at all. There should either be a super-agent that orchestrates and does no work, or a long running agent that can retrieve from previous sessions, not a worker that completes tasks partially and delegates the rest to other workers
Don't design harnesses based on our assumptions of human behavior
Yaser Martinez retweeted
when bend2 launched, some complained about AI usage. in my view, this is only a problem if bugs can occur. the current typescript compiler, for example, has many. I hate it. with well-designed laws, though, that ceases to be a problem. so, for the better or for the worse, bend3 will be 100% AI-written. there is no going back
while I was raising, I left a fleet of agents working on bend3, a small scale version of this vision. this was a huge success, beyond my own expectations. opus 5.5 successfully rewrote all of bend in itself, including its compiler, which is now certified, with zero bugs, and already 2x faster (!!). we now have a model of Bend, C, CUDA and the compiler, all written in Bend itself, and properly formalized. with this, we'll soon start the largest proof-driven compiler self-improvement loop in history
news on Bend have been slow because I was busy raising, but as soon as this is over, expect updates at a breath-taking place. this is the most exciting thing I've ever done and I can't wait for the big run to start
Yaser Martinez retweeted
what if code mode ran *inside* the sandbox?
today's code mode cannot:
- write reusable scripts that call tools
- import packages (npm / pypi)
- read & write files too large to fit in memory
- call tools after turn ends
so here's a prototype of something new [demo @ 1m30s]
~~~
what is the point of code mode in the first place?
agents love calling tools via code. it lets it:
- chain tools together
- discover tools without bloating context
- typecheck scripts
- more efficient token usage
code mode also simplifies auth by running tools directly in your trusted backend context
~~~
so what's the problem?
what's running inside your sandbox (node/python/bash) cannot access your agent's tools
it separates the tool calls from the context of the sandbox:
- cannot write reusable scripts that call tools
- cannot use rich npm / pypi packages
- cannot work with large files larger than memory
- cannot call tools while agent is idle
~~~
so how do you fix this?
your agent needs bi-directional tool calling
your agent can execute bash inside the sandbox
β i.e. call scripts
and your sandbox can execute tools on your agent (new)
β i.e. execute mcp
~~~
how do you implement this?
the simplest way is to:
- allow your sandbox to call your agent tools over HTTP
- secure those tools with a JWT (scoped to sandbox)
it's similar to auth proxies that inject tokens, but for your tools
~~~
what does this look like to the agent?
expose your tools as a script ππππ-ππππ.ππ
when the agent needs to invoke code mode, node.js can call your tools:
ππππ -π "ππ‘ππ('πππππ-ππππ.ππ πππππππ')"
(contrived example)
this allows code mode to have full access to your sandbox with:
- npm / pypi packages
- write reusable scripts and execute them
- work with files on the system that don't fit in memory
- generally better ergonomics for the agent
~~~
show me the code
i attached working example code & skill below
it's using Pi Durable on Rivet Actors
Rivet comes with JWT built in, which makes it really easy to expose its tools to the sandbox
Yaser Martinez retweeted
Rivet is the open-source alternative to Durable Objects
And it is fast:
39Γ faster β actor creation
28Γ faster β SQLite writes
13Γ faster β request from your backend
Yaser Martinez retweeted
1/4
Meet Qwenputer: Qwen3.5-4B with a full RISC-V CPU inside its weights, using the model's own Gated DeltaNet layers as RAM.
It runs C and MicroPython in its forward passes, no external runtime.
On DeepMind math: base 55β90%, base + Python tool 95β100%, Qwenputer 97.5β100%.
Yaser Martinez retweeted
1/
Hiring your first employee is one of the biggest moments for a founder. But payroll, taxes, benefits, and compliance are far more complicated than they need to be.
Now with AI we can make the pain of payroll go away.
@Mercury Payroll* launches today!
Yaser Martinez retweeted
We took NVIDIA's Parakeet and made it better.
Introducing Fluid Speech Mini: the best Open Source Model purpose-built for Dictation π
With SOTA performance under 300 MB
Run it on your iPhone, Mac, or PC. No cloud, no internet.
Dictation shouldn't need a server.
Your voice should be yours and we will make sure of that.
Open weights and more info below!
Yaser Martinez retweeted
Foldkit 0.166.0 is out!
A new experimental Query module provides reusable Submodels for fetching, caching, and refreshing remote data, with less boilerplate.
Plus Effect 4 stable (in 0.165.0) and other improvements.
foldkit.dev/blog/foldkit-0-1β¦
Yaser Martinez retweeted
Here are the slides from my internal tech talk at Scalar: a survey of WAL-on-S3 database architectures. Feedback is welcome!
choplin.dev/slides/wal-on-s3β¦
Yaser Martinez retweeted
S3 is the only dependency in Rivet 3.0 (coming soon)
It's architected for easy self-hosting and embedding in your own apps
But that also meant getting rid of NATS....
Yaser Martinez retweeted
If youβre younger than 45, this is the first time in your life that youβre experiencing a real rising interest rate environment.
The last time this happened was 1960 - 1980.
This fundamentally changes the physics of the game in a way many wonβt be prepared for.
Yaser Martinez retweeted
in life there is no free dopamine lunch, except bromantane.
this stuff releases dopamine. inhibits its reuptake AND then replenishes your dopamine stores once they are depleted.
dopamine is the chad neurotransmitter: ambition, drive, focus, motivation, willpower and the relentless urge to fucking win. bromantane literally makes you a winner.
did i mention it is also anxiolytic despite being a literal stimulant?
Yaser Martinez retweeted
1/ We're reimagining what it means for your agents to work with all your computers.
Today we're excited to share Cua Spaces, built on Cua Driver and our virtualization stack, rolling out on macOS today.
Download it now, it's free and source-available: spaces.cua.ai
Yaser Martinez retweeted
@supabase has acquired @tursodatabase
i'm been a huge fan of the team and what they've built. we have big plans together. more to share in our keynote in a few hours
this is the tweet that convinced us they'll fit right in
xitter.cf/glcst/status/209307402β¦
Yaser Martinez retweeted
Another insane Jev use case!
Jev makes it incredibly cheap to evaluate and classify agent runs at scale.
And finally, someone open-sourced a self-improving memory layer that can put that capability to work across agent harnesses.
It turns your agent sessions into a compounding knowledge layer, where every successful run can make future agents smarter across:
- Codex
- Claude Code
- Cursor
- OpenCode and 20+ more
Beacon by @asymptotelabs continuously builds a shared history across your agent harnesses and uses Jev to identify the runs worth learning from.
It then turns the best workflows, corrections, and debugging patterns into reusable skills.
GitHub repo: github.com/Asymptote-Labs/agβ¦.
(donβt forget to star it β)
Most agent runs are messy.
They contain exploration, failed commands, dead ends, and one-off fixes that should never become permanent memory.
So Beacon preserves the full session history, while Jev helps decide what should be promoted, reviewed, or discarded.
The recording below shows this in action.
Beacon found 579 sessions across 5 coding-agent harnesses and normalized them into one consistent history.
From there, Jev surfaces the lessons worth keeping and makes them available across your agent stack.
- A pattern learned in Cursor can carry into OpenCode.
- A lesson from Claude Code can improve the next Codex run.
Every successful run adds to the shared knowledge layer, making future agents smarter.
If you want to dive deeper into Jev, I also wrote a breakdown of how it works.
The article is quoted below.
