I’m building Orglet, a desktop app where you give AI workers (I call them orglets, and they are cute) their own instructions and talk to them individually or together in a group chat.

You can build your own army of orglets using your existing Claude Code or Codex subscription, an API provider, or even locally hosted models.

Like other agent harnesses, orglets can access a working folder you give them permission to use. They can split the work among themselves, and you can review the resulting diff before applying changes to your files. Each orglet has its own role, tools and permissions.

Orglet is free under AGPLv3. Downloads are available for Windows, macOS and Linux. It is still very early and flaky, so I’d appreciate any contributions. Thank you!

I have a lot of experience with LLM’s and have seen them become very good at interacting with tools within an environment. I’m truly amazed at how poorly the new Siri is at handling very basic requests for basic interactions with different parts of the ecosystem. I’m really beginning to question whether LLM technology is being leveraged at all here. What am I missing with this? I have tons of examples where it fumbles softball requests.

For example: “Siri, can you give me the location of my driver. He just sent me a text message with his location.”

Response: Siri shows me a text from 5 days ago about a Seamless driver. Instead of a message just received that says “I’m at terminal 2 door B.”

This is beyond table stakes. In the message, I’m giving some really easy guidance on which tools to use and how to use them. For example, I mention text messages so we already know that we should be searching messages in the Message app. And if I say just messages me, it should be looking at the most recent texts. Not from 5 days ago.

And I have a whole slew of other examples that the old Siri was able to nail perfectly.

Example: “Make an appointment for Monday from 12 to 4 and label it Furniture Delivery.”

Response: Siri added an appointment for 3:48 to 4:48 today.

Then I said, “No, from 12-4.” It then moved another appointment to 4.

Yet Another Example: “Siri, send Jane a message saying I’m on my way home.”

Response: “Sure, what do you want the message to say?” (I just told you!)

What am I missing? A basic LLM can handle this. Why is this so poor? I’m not understanding why, if Gemini is the backend, it can’t handle this better.