Alaeddine Abdessalem

AI Engineer

[DeepSeek @deepseek_ai]

🚀 DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length.

🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models.
🔹 DeepSeek-V4-Flash: 284B total / 13B active params. Your fast, efficient, and economical choice.

Try it now at https://t.co/GCdiMzk1Dl via Expert Mode / Instant Mode. API is updated & available today!

📄 Tech Report: https://t.co/drlDrxkYtp
🤗 Open Weights: https://t.co/T13Y8i7SDM

1/n

[Alaeddine Abdessalem @alaeddine_abd]

We’re moving away from typical AI Agents (prompt+tools) to Agent Harnesses, extensible in capabilities and memory. Claude built the infra to integrate a harness like claude code into production apps, but no one is yet building something open source, for other harnesses like Hermes agent of OpenClaw.
I’m trying to build that for python developers. pip install harnesdk

[Claude @claudeai]

Introducing Claude Managed Agents: everything you need to build and deploy agents at scale.

It pairs an agent harness tuned for performance with production infrastructure, so you can go from prototype to launch in days.

Now in public beta on the Claude Platform.

[Alaeddine Abdessalem @alaeddine_abd]

Repo link: https://t.co/9P5oPUKfvr

[Alaeddine Abdessalem @alaeddine_abd]

If you're still building your own agent harness, you're not moving at full speed.

Developers are picking an existing harness like Claude Code or OpenClaw and extending it with skills to build vertical AI products fast:

- UGC creation with 𝘳𝘦𝘦𝘭𝘤𝘭𝘢𝘸 skills
- GTM engineering and lead generation with 𝘨𝘵𝘮-𝘱𝘰𝘴𝘪𝘵𝘪𝘰𝘯𝘪𝘯𝘨-𝘴𝘵𝘳𝘢𝘵𝘦𝘨𝘺 skills
- animations with 𝘳𝘦𝘮𝘰𝘵𝘪𝘰𝘯 skills

The right harness + the right skills = a powerful vertical AI app. The pattern is real and it's accelerating.

But integrating a harness comes with a hidden infrastructure burden. An agent must live inside a sandbox, not on production infra. Isolated, ephemeral, safe.

Solutions like E2B offer this critical layer, yet you still need to integrate it with the harness.

I built 𝗛𝗮𝗿𝗻𝗲𝗦𝗗𝗞 to close this gap.
A minimal API, that gives you a customizable agent harness like Claude Code living inside a sandbox.

`pip install harnesdk`

If you're building AI agents and tired of reinventing the infrastructure layer, give it a try.
⭐ the repo and run your first agent in under 5 minutes. 👇

[Alaeddine Abdessalem @alaeddine_abd]

Link to the google colab notebook: https://t.co/odk8tsO7cM

[Alaeddine Abdessalem @alaeddine_abd]

Gemma 4 is here, I used it to run openclaw or hermes agent on free GPU hardware!
Gemma 4 was released days ago, and according to @ArtificialAnlys its intelligence is comparable to models 10x its size (see attached image).

The edge models (e2b and e24) can be run on free T4 GPUs on google colab.
This means, you no longer have to buy a mac mini to install OpenClaw or @NousResearch Hermes agent and connect it to Anthropic API to run it.
Running Hermes or OpenClaw can be quite expensive if it's connected to Anthropic API especially if you have long running agents or background tasks.

Here's how I setup on google colab:
1. Connect your colab notebook to T4 GPU
2. Install ollama and run a background ollama server
3. pull gemma4:e2b or gemma4:e4b model
4. Install and configure Hermes to use the locally running Gemma model

> Long running Hermes agent for free 🆓💸

If open source AI models become smaller, more efficient and runnable on commodity hardware, are billions of AI infrastructure investments still justified ? Will future efficient models burst a potential AI bubble ?

Colab link in 🧵

[Alaeddine Abdessalem @alaeddine_abd]

Check the repo here: https://t.co/4EFXNc0kIC
It’s getting almost 50k stars in one day

[Alaeddine Abdessalem @alaeddine_abd]

If you’re building agent harnesses, the recently leaked Claude Code codebase is a gem.
I’ve been curious about how Claude Code is implemented. Here are the answers to some questions I long had after some digging (using claude code itself):

1. What’s unique about its system prompt?

It’s assembled dynamically and split into two tiers : a static, globally cacheable layer (tone, safety rules, task philosophy) and a dynamic, session-specific layer (memory, env info, MCP tool descriptions, token budgets).

2. What tools does Claude Code use?

35+ tools across file I/O, search, shell execution, web, task management, scheduling, MCP integration, and agent coordination. Notable ones: LSPTool for live IDE diagnostics,
and TeamCreateTool for spawning in-process agent swarms.

3. How is the file edit tool implemented?

Looks like a typical string replacement with some normalisation logic.
Edits are rejected if the target string appears more than once, if the file wasn’t read first in the session, or if the file changed between read and write. File edits trigger LSP by default.

4. How is context compaction implemented?

A three-layer system running every turn. First, it trims old tool results (microcompact). Then it tries fast-path session memory. If that fails, it runs a full summarization into 9 structured sections. There’s a circuit breaker after 3 consecutive failures.

5. How does the subagent system work?

The subagent system lets Claude Code spawn, coordinate, and communicate with child agents. Agents can run in-process, as background tasks, or on remote infrastructure.
Swarm teammates share the Node.js process via AsyncLocalStorage and communicate through a file-based mailbox.

6. How are skills implemented?

Skills are named prompt workflows backed by .md files or compiled TypeScript. They run either inline (injected into the current conversation) or forked (isolated sub-agent with its own token budget). At each turn, available skills are listed in a <system-reminder> block capped at 1% of the context window.

There’s a lot more to learn from the source code, like the memory system, mitigating context bloat, planning, plugins,…

link to the repository of the leaked codebase in 🧵

[Daniel Hnyk @hnykda]

LiteLLM HAS BEEN COMPROMISED, DO NOT UPDATE. We just discovered that LiteLLM pypi release 1.82.8. It has been compromised, it contains litellm_init.pth with base64 encoded instructions to send all the credentials it can find to remote server + self-replicate. link below

[Alaeddine Abdessalem @alaeddine_abd]

@mmabrouk_ @ivanburazin I feel like APIs are kind of an imperative language. Maybe project as code is a better way to interact with software? For instance defining cloud infrastructure as code might be more intuitive for LLMs than a list of API calls to make.

[Alaeddine Abdessalem @alaeddine_abd]

@andreasklinger Amazing video! I see why you said SaaS is over and suggested to move into robotics. I’m wondering whether a degree is necessary for such a transition from software

[Alaeddine Abdessalem @alaeddine_abd]

@SuperDesignDev Check https://t.co/09LOiYgJmD for more styles

[Alaeddine Abdessalem @alaeddine_abd]

"Introducing Styles2.0" - Win98 style
We integrated new design styles from @SuperDesignDev in Hera!
Check the styles page to learn how to one shot stunning animations like this one

[Alaeddine Abdessalem @alaeddine_abd]

@Kimi_Moonshot Congrat!
Looking forward to try this for video animations!

[Alaeddine Abdessalem @alaeddine_abd]

Figma to animation with a simple prompt

[Peter Tribelhorn @tribelhorn23]

You can now animate Figma designs in Hera with 1 click.

[Alaeddine Abdessalem @alaeddine_abd]

3D animation

[Alaeddine Abdessalem @alaeddine_abd]

I made this @heravideoapp logo animation in under 2 minutes 🤯
This would have cost hundreds of dollars if made by a professional

[Alaeddine Abdessalem @alaeddine_abd]

More styles in the thread

[Alaeddine Abdessalem @alaeddine_abd]

Vibe editing is the way forward!
I joined @heravideoapp with the mission of making motion design accessible.
If you want to create animations without installing any tool and with manual edits, try out Hera!

[Remotion @Remotion]

Remotion now has Agent Skills - make videos just with Claude Code!

$ npx skills add remotion-dev/skills

This animation was created just by prompting 👇