@astralwave Can you break it down how you go from coarse ask to small parts? What’s the story behind your screenshot?
English
Josselin Perrus
2.8K posts
Codex sub-agents are such a treat to use
I love how you can steer each one individually
Makes it possible to start with a coarse ask, have the agent break up the work into small orthogonal parts, then deep dive into each one
It’s also super cute that each sub-agent gets a name like “Erdos”, “Halley”, “Hubble”, etc: it’s a small detail but it makes it possible to refer to them easily!
English
SAM-3 provides remarkably precise answers (0.97 IoU).
But 1 out of 5 times it returns nothing.
I suspect it to be a prompting issue.
SAM-3 is currently a solid addition to coding agents. Will the general models catch up?
Full write-up: meaningfool.net/articles/detec…
English
DINO alone scored worse than raw LLMs, regularly choosing the wrong shape (triangle instead of rectangle), but with high precision
The target was often in its top-5 candidates though.
Delegating the ranking to Claude Code with a Skill improved the results substantially.
English
Agents/LLMs can't locate what they see on images.
They can describe an image, its composition. But they can't "select" an element to move or resize programmatically: the bounding box they produce is off.
Can tools fix that? I ran a small experiment.
English
Josselin Perrus retweetledi
New paper on skills. The conclusions hold up exactly 1-to-1 with our experience.
Skills are better than docs, but only when made with care. Less is more.
Models are really bad at making skills. 2 paragraphs of human-written condensed instructions (or best practices) are better than 20 pages of model-summarised info. My theory is they lack the theory-of-mind IQ to think through what they might have needed to know before they knew it.
English
@masondrxy Fixed on mobile. I mention it, but did not go much further as i did not dig into it yet.
English
@nonils all good! good content.
would love to see Deep Agents mentioned under Agent SDKs! open alternative to Claude Agent SDK:
docs.langchain.com/oss/python/dee…
English
Agent frameworks are confusing. So I wrote "Agent Frameworks for the Rest of Us" as I dove into the topic.
Turns out you can classify them in 3 categories: "Orchestration Frameworks", "Agent SDKs" and "Agent Servers"
👇
English
@nonils can’t view anything on mobile :( sidebar covers everything
English
@NathanFlurry I featured sandbox-agent in my "agent frameworks for the rest of us" report.
It's by a non-developer (me), for non-developers. Hopefully I got things right.
Would love to have your feedback
x.com/twitter/status…
Josselin Perrus@nonils
Agent frameworks are confusing. So I wrote "Agent Frameworks for the Rest of Us" as I dove into the topic. Turns out you can classify them in 3 categories: "Orchestration Frameworks", "Agent SDKs" and "Agent Servers" 👇
English
"a lot of names for a lot of things"
Cloudflare has 9 product names for what is essentially Durable Objects under the hood
Rivet Actors just give you the primitive as a lightweight library. No vendor lock-in and no confusing marketing.
simeonGriggs@simeonGriggs
Cloudflare has a lot of names for a lot of things—finally took the time to figure (and draw) some of them out making this real-time app.
English
@modal @tryramp @modal @charles_irl I wrote a report about agent frameworks "for the rest of us" - the non-dev, as I am.
I report about the Ramp / Modal architecture, among others.
If you give it a look, let me know what you think
x.com/nonils/status/…
Josselin Perrus@nonils
Agent frameworks are confusing. So I wrote "Agent Frameworks for the Rest of Us" as I dove into the topic. Turns out you can classify them in 3 categories: "Orchestration Frameworks", "Agent SDKs" and "Agent Servers" 👇
English
@threepointone @threepointone I published a report on agent frameworks. There is some @cloudflare inside. But no Cloudflare Agent SDK as I'm not yet familiar with it
Do you think it would fit in the model I define like the PydanticAI, Vercel AI SDK, Pi and Claude Agent?
x.com/nonils/status/…
Josselin Perrus@nonils
Agent frameworks are confusing. So I wrote "Agent Frameworks for the Rest of Us" as I dove into the topic. Turns out you can classify them in 3 categories: "Orchestration Frameworks", "Agent SDKs" and "Agent Servers" 👇
English
…-playground.threepointone.workers.dev (beta)
a demo of every agents sdk feature. it's a lot.
we're lacking features from other frameworks/infra, and we'll close that gap very quickly. but there are features in our agents sdk that will never be in other frameworks/infra.
English
Gemini 3.1 pro is a GREAT animator
LLM was struggling with spatial positioning when generating animation before,
But look at those Gemini 3 pro VS Gemini 3.1 pro comparison
Gemini 3.1 pro visibly reasons better
You can try it free on @SuperDesignDev today
----
I might also share some prompt guide on getting most out of model for animations, comment below if you interested
English
@steveruizok an interaction we had on LinkedIn about your project to build an internal agent for @tldraw actually triggered my effort to better understand Agent Frameworks.
And that's me sharing what I learnt. Thanks for the involuntary nudge :)
x.com/nonils/status/…
Josselin Perrus@nonils
Agent frameworks are confusing. So I wrote "Agent Frameworks for the Rest of Us" as I dove into the topic. Turns out you can classify them in 3 categories: "Orchestration Frameworks", "Agent SDKs" and "Agent Servers" 👇
English
@badlogicgames @thdxr @samuelcolvin I was curious about how Pi, Opencode, PydanticAI and other framewoks were different. So I wrote about what I learnt.
I'm not a dev though so hopefully there aren't too many inaccuracies. Feedback welcome :)
x.com/nonils/status/…
Josselin Perrus@nonils
Agent frameworks are confusing. So I wrote "Agent Frameworks for the Rest of Us" as I dove into the topic. Turns out you can classify them in 3 categories: "Orchestration Frameworks", "Agent SDKs" and "Agent Servers" 👇
English
@JacobColling your experimentations with Claude Agent SDK picked my curiosity. Being less technical, I was a bit confused about all those agent frameworks out there.
I wrote about what I learnt diving into the topic. Would love to have your review
x.com/nonils/status/…
Josselin Perrus@nonils
Agent frameworks are confusing. So I wrote "Agent Frameworks for the Rest of Us" as I dove into the topic. Turns out you can classify them in 3 categories: "Orchestration Frameworks", "Agent SDKs" and "Agent Servers" 👇
English
If you are wondering what the difference is between LangGraph, PydanticAI, Claude Agent SDK, Vercel AI SDK, Opencode... So did I.
The report is probably not 100% accurate, but I believe it can help. Feedback welcome.
…ent-frameworks-report.meaningfool.net
English
In this report, I cover:
- How we got from prompting to agents
- App-driven vs agent-driven control flow for orchestration
- Why access to bash and filesystem is a game changer
- Architecture implications of long-running agents
- Analysis of 4 real projects
English