HieuRistic

[Vox @Voxyz_ai]

A lot of people have asked how to configure this cost-efficient Astra + Luna + Sol agent tree and get it working.

There are a few ways to use it, depending on how you normally use Codex.

1. Have Codex configure it

Give Codex the diagram and ask it to set up project-level custom agents. Specify the model and model_reasoning_effort for each role, and save the rules for assigning work. Once the configuration is loaded in a fresh session, Astra can call subagents using those roles.

2. Turn it into a Skill

If you use this setup often, put the role configuration templates into a Skill, along with when to use each role and how to collect its results. Set it up the first time, then use it as needed without explaining the whole tree every time.

3. Use it through Herdr

If you already use Herdr, you can launch and manage Codex CLI there. Herdr manages the sessions, while Astra assigns work within Codex using the same role configuration. No need to change the setup just to use Herdr.

Whichever approach you use, make sure the role configuration is active. Writing “have Luna explore and Sol implement” in chat or AGENTS.md doesn’t mean the model has switched. You can ask Codex to check the subagent session logs to confirm the model and reasoning effort.

If you just want to get started, the first option is enough. Small tasks don’t need the whole tree, either. Of course, everyone’s experience with quota usage across models is different, so you can adjust the models and effort levels to suit your own usage.

[Prajwal Tomar @PrajwalTomar_]

In last 5 months, I’ve built 16 SaaS products for clients using Cursor.

Now, I’ve cracked the best AI coding workflow for Cursor.

Here’s my step-by-step guide to building production-ready MVPs:

[Assem @achammah1]

We built the most versatile and powerful Agent-RAG integration with nexusGPT new Knowledge Update:

- Multi-modal embedding : Automatically treats document with images /figures /charts /illustrations /schematics and can perform RAG over their content (inspired by the work of @hrishioa)

- Multi-level abstraction : Automatically creates trees and multi-level summaries over specified set of documents to enable holistic understanding of knowledge base (e.g., asking abstract questions about documents that are not necessarily present into a specific subset of chunks) (inspired by Inspired by the impressive work of @parthsarthi03’s RAPTOR algorithm and cookbook in @hwchase17’s @langchain)

- Chunk control : Decide on how you want to embed your document (by default Semantic Chunking invented by @GregKamradt)

- Document type: Not only can be regular documents (pdf, docx, pptx, txt, using @UnstructuredIO) but also @NotionHQ and any Websites (automatically fetches sitemap)

- Knowledge Tool : Integrate your knowledge and create specialised RAG tool that your agent can call given user query (inspired by a post from @yoheinakajima)

- Knowledge Context: Improve your agent prompt with specialised RAG variable over your data to enhance every query sent to your agent

On top of that, you can now integrate your agent on @SlackHQ and @WhatsApp, SMS, @facebook Messenger (through @twilio)- Oh and we also vastly improved the chat experience, you can now seamlessly switch between any agent/model by just @'ing them