Bellverse Labs

An agentic creative studio.

[anabology @anabology]

Open source suno soon, wow

[World Labs @theworldlabs]

We’re opening up access to Atlas to everyone today.

We’ve built apps and API endpoints to help you use Atlas any way you can imagine: bring your ideas for building a 3D world and come away with keyframes, video clips, or 3D Gaussian splats.

[きしだൠ(K1S) @kis]

Fable 5.2にクラシックギターの物理シミュレーションやってもらった。ちゃんと鉄弦とナイロン弦が違う。
ボサノバギターソロってやったらちゃんとボサノバになってる。ただ、人間に弾ける運指ではない・・・

[Vercel @vercel]

Introducing the UI SDK.

[Chris Albon @chrisalbon]

Don’t show me something you did with one prompt. Show me something you did over three months with 1000 prompts, five resets, and a dozen all nighters.

[OpenAI @OpenAI]

Now in beta: Businesses in ChatGPT.

Claim your business, manage your information, and get insights on how people engage with you.

Eligible businesses can also build a Business Agent to answer customers’ questions and help them explore what you offer.

https://t.co/N2GEZX0mFT

[Bellverse Labs @BellverseLabs]

billion-context: a compression plugin for Claude Code, Codex and Hermes that claims month-long sessions on a 100K window at 5x fewer tokens. https://t.co/V7ISbLJ66Q

[Bellverse Labs @BellverseLabs]

Would you rather run a personal agent locally in a VM, or on a server that keeps working when your laptop sleeps?

[Paul Iusztin @pauliusztin_]

I've FINALLY wrapped my FREE, open-source course on building a coding agent from scratch!

• 8 in-depth articles.
• 6 video lessons.
• 1 complete roadmap to mastering harness engineering.

Over the past few months, I've studied the internals of Claude Code, OpenCode, Pi, and Aider to understand what makes coding agents work.

And I noticed the agent loop is the easiest part.

The actual engineering happens in the harness.

Things like:

• Context management
• Sandboxing
• Permissions
• Memory
• Tool execution
• Evaluations

So I've turned everything we learned into a practical course where you build a fully functional Claude Code clone.

Here's the complete 8-step roadmap:

1. Harness Architecture: Design the core agent system and its components.
2. Agent Loop: Build the reasoning and tool execution loop from scratch.
3. Sandboxing: Execute code safely using Docker and Modal.
4. Context Engineering: Implement memory, skills, compaction, and LSP.
5. Subagents: Run parallel agents with isolated context windows.
6. Remote Execution: Deploy agents to the cloud using CLI, webhooks, and cron jobs.
7. Evaluations: Build benchmarks and regression tests with Opik.
8. Recording & Replay: Reproduce failures and verify fixes with Kitaru.

By the end, you'll have a working coding agent with an interactive terminal interface AND a headless runtime for running agents remotely.

You can even try the finished agent before starting the lessons.

Everything is written in Python.
And the entire codebase is open source under Apache-2.0.

A HUGE thank you to @modal, @Cometml, and @zenml_io for sponsoring this course and helping us keep it 100% FREE.

Whether you're an AI engineer or a software engineer transitioning into AI, this is the roadmap I wish I'd had when learning harness engineering.

Get the complete course on GitHub: https://t.co/JmCajfijIP

And follow along with the articles on Substack: https://t.co/MLZHZRdRmd

P.S. If you find it useful, give the repo a star so more engineers can discover it.

[Bellverse Labs @BellverseLabs]

what do you check before you trust your coding agent when it says it's done?

[klöss @kloss_xyz]

Fable 5.2’s advice for the next 10 years.

[Adam Holter @AdamHoltererer]

GPT-4 level intelligence has gotten roughly 1800x smaller

[Guillermo Rauch @rauchg]

Built a physical simulation of a record player with WebGL2 and Web Audio. It's *shockingly* accurate: https://t.co/MBNaTfKoai.

Instead of playing a song, it works by converting it into the physical shape of a groove. Then a simulated stylus rides those walls 384,000 times a second, and you hear the voltage that motion would induce in a cartridge.

So the crackle, the pops, the wow/flutter, the dull inner grooves are natural byproducts of the simulation. You can also "DJ" the record.

(The converted songs are in the public domain or CC-licensed. Code is linked for you to remix and re-deploy.)

[Bellverse Labs @BellverseLabs]

on a 64gb box, would you rather run a huge model quantized way down, or a smaller one at higher precision?

[Bellverse Labs @BellverseLabs]

DwarfStar from @antirez now runs DeepSeek V4.1 Flash with SSD streaming.

The kimi-k3 branch also runs Kimi K3.1 Pro from safetensors. Slow, but good enough for QA tasks.

https://t.co/agdgadbi06

[undertow @undertow_tez]

what have I done?! 😂😭

[Bellverse Labs @BellverseLabs]

How does real-browser MCP compare to a headless tool for the tasks you actually assign agents?

[Bellverse Labs @BellverseLabs]

devvm gives each project folder a Lima VM with the toolchain detected from your files. Claude Code, Codex and OpenCode run inside it; only the project is mounted, network defaults to an allowlist, and secrets stay on your machine. https://t.co/62my7utK62

[Bellverse Labs @BellverseLabs]

OpenCode Browse is an MCP server plus Chromium extension so coding agents drive your real browser—tabs, logins, extensions—with handoffs for 2FA and payments. Set a site allowlist before the first run; HAR captures can hold cookies. https://t.co/bsjdAWYf3A

[Chris Tate @ctatedev]

Introducing @𝚓𝚜𝚘𝚗-𝚛𝚎𝚗𝚍𝚎𝚛/𝚌𝚑𝚊𝚝

Generative UI for Slack, Teams and more

→ Agents reply with native cards
→ Tables, fields, charts and buttons
→ Streams in as the model writes
→ One catalog for every Chat SDK platform