Nabil Belakbir

[Nabil Belakbir @nabilblk]

I'll publish more scenario logs. New agents are joining the board, so work is resuming . the tasks here are also meant to give them awareness of the current status. I chose a minimalistic set of concepts: I'll add artifacts( agent outcome ) and budget as first-class concepts and see how that affects swarm efficiency. An eval dataset for coordination tasks would also be a good direction to build.

[Nabil Belakbir @nabilblk]

I’m open-sourcing Harakiri Blackboard: a shared coordination layer experiment for independent Claude Code and Codex agents, with humans in control.

Coordination deserves its own infrastructure, outside the harness. https://t.co/EgCtUdbk8f

[Nabil Belakbir @nabilblk]

@shlokbuilds 1 blackboard , multiple agent working on it , coordination is managed by an Agent asigned as coordinator. So Multiple mixed runtime . More runtime support will be added in the next days (Grok , opencode ....)

[Nabil Belakbir @nabilblk]

Agreed. A tool result shouldn’t become a trusted instruction just because an agent posts it. The coordination layer needs explicit authority, provenance, and a clear distinction between evidence and instructions. "Results" or "Outcome" could become also a first class concept . (Next release 😀)

[Nabil Belakbir @nabilblk]

@insidersweb3 Exactly. The board should preserve context, decisions and ownership as agents come and go and let them adapt to new events, reorganize work and try new coordination patterns. That’s why coordination needs its own observable layer outside the runtime/harness.

[Nabil Belakbir @nabilblk]

@alohaproxy Thanks for the suggestion! I’m keen to get Blackboard in front of people running real multi-agent setups. I will deep dive on HopUp.

[Nabil Belakbir @nabilblk]

@mianoedwin_ Exactly. Coordination rules and constraints should live outside the harness, where we can enforce and observe them.

I’m not measuring handoffs systematically yet (Still fresh :) ). I want to track both handoff latency and failed or duplicated work , the cost shows up in both.

[Nabil Belakbir @nabilblk]

@addyosmani Its time to say enough is enough .feel tired , LOC isn t a metric . This is o1o in swe. Your leaked code show how worst and bad code you produce and now you insist on communication . Go and refactor

[Nabil Belakbir @nabilblk]

𝗥𝘂𝗻𝗻𝗶𝗻𝗴 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 𝗼𝗻-𝗽𝗿𝗲𝗺𝗶𝘀𝗲𝘀 𝗶𝘀 𝗼𝗻𝗹𝘆 𝗽𝗮𝗿𝘁 𝗼𝗳 𝘁𝗵𝗲 𝘀𝗼𝘃𝗲𝗿𝗲𝗶𝗴𝗻𝘁𝘆 𝘀𝘁𝗼𝗿𝘆.

You also need control over where its actions happen.

While building a sovereign, on-premises cybersecurity platform, I found myself working on exactly that problem. Both defensive investigations and offensive security testing call for capable models, rich and sometimes atypical tool usage, and potentially models whose safeguards or alignment cannot simply be assumed.

An agent needs somewhere to execute code, install dependencies, manipulate files, and run tools. But putting it in a sandbox is only the beginning.

• Who can launch those environments?
• What can they access?
• How do you inspect a task after disconnecting?
• Which files should survive when the runtime stops?

And how do you make all of this straightforward for the developers integrating it?

𝗔𝗴𝗲𝗻𝘁𝘀 𝗻𝗲𝗲𝗱 𝗮 𝗵𝗼𝗺𝗲. But as they become more capable, that home increasingly needs the boundaries of a jail: controlled access, explicit permissions, and limits enforced outside the agent itself. I believe that is the right design choice.

Those questions became 𝗛𝗮𝗿𝗮𝗸𝗶𝗿𝗶 𝗦𝗮𝗻𝗱𝗯𝗼𝘅: a self-hosted sandbox control plane for agent applications.

Sandboxing is an established field. Daytona, E2B, and Vercel Sandbox are already doing substantial work here.

My focus is the experience around execution: giving teams a consistent way to prepare environments, manage access, run work, inspect results, retain working files, and clean up.

The dashboard, CLI, and TypeScript SDK expose the same control-plane concepts. Your application continues to own the agent and its orchestration.

A sincere thank-you to the 𝗢𝗽𝗲𝗻𝗦𝗮𝗻𝗱𝗯𝗼𝘅 𝗺𝗮𝗶𝗻𝘁𝗮𝗶𝗻𝗲𝗿𝘀. Their runtime provides the execution foundations Harakiri Sandbox uses today. Execution sits behind a provider interface; the control plane is where I’m focusing this project.

I’m sharing Harakiri Sandbox under 𝗔𝗽𝗮𝗰𝗵𝗲-𝟮.𝟬 because I want the work to be useful beyond my own project.

AI-assisted development is changing how we build and adapt software. My approach to sharing is straightforward: use it, study it, contribute, or take a fork in a direction I haven’t considered. Contributions are welcome, but they aren’t the only valuable outcome.

This is a 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿 𝗣𝗿𝗲𝘃𝗶𝗲𝘄, with documented capabilities and limitations. Self-hosting is a deployment choice, not a security guarantee.

The video shows the actual dashboard and a complete workflow.

If you’re building an agent product, try it with one real task. I’d particularly value feedback on where integration is still harder than it should be.

𝗩𝗶𝘀𝗶𝗼𝗻 𝗮𝗻𝗱 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲:
https://t.co/S5xHUQ67Rg

𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻:
https://t.co/hMglo4qCsb

𝗣𝗿𝗼𝗱𝘂𝗰𝘁 𝘁𝗼𝘂𝗿:
https://t.co/dKcRFPs6G2

𝗥𝗲𝗽𝗼𝘀𝗶𝘁𝗼𝗿𝘆:
https://t.co/fHCTnNn5HA

[Nabil Belakbir @nabilblk]

Focusing on cheating and plagiarism is the wrong angle.

Remember the iconic battles: Newton vs. Leibniz, Einstein vs. Hilbert.

We finally land… confused.

[Nabil Belakbir @nabilblk]

Crazy times , Officially Grigori Perelman is the first and last human to resolve a Millennium Prize Problem. And suddenly i remember Lee Sedol pain after game 3 saying he was sorry .

[Nabil Belakbir @nabilblk]

The most interesting part isn’t just the multimodal interaction model, but the split architecture: a real-time model coordinating with a background agent. Feels less like a pure model capability and more like model system codesign at inference time. Any chance this can be a post-training Topic ?

[Nabil Belakbir @nabilblk]

@rezoundous Build your eval and back to real swe again

[Nabil Belakbir @nabilblk]

@aiedge_ Look at the leaked Claude code, the troubles, regressions, recent incidents, and the highly expensive troubleshooting efforts , they are all related to the poor software engineering principles applied

[Nabil Belakbir @nabilblk]

@Samaytwt Can he lose his job due to AI?

[Nabil Belakbir @nabilblk]

@DanielMiessler Jensen just highlighted a big US misconception. He pointed to the telco story nobody wants to see: Huawei is already the king.
The same thing can happen now , GLM 5.1 was fully trained on Huawei Ascend. That’s exactly what he meant by "they already have the ship". 2nm is a detail

[Nabil Belakbir @nabilblk]

@trq212 So you are overfitting .

[Nabil Belakbir @nabilblk]

@trq212 @theo You should roll back and fix the issues , there are a lot of them. I switched to a pure dummy model (we are too far from automating software engineering).

[Nabil Belakbir @nabilblk]

@bcherny @GergelyOrosz Please explain where this “enterprise demand” came from? I can’t conceptualize a company asking to pay 10x more.
Be transparent , your credibility is on the table.

[Nabil Belakbir @nabilblk]

@_afadil @WealthfolioApp Very cool