I built my own agent framework and it's awesome
retained text
**The BizAgent Framework** It started out as a system prompt for Claude Code and eventually evolved into what I call BizAgent (in hindsight a horrible name because GitHub has about a thousand projects with that name lol), a complete agentic hub-and-spoke system. The basic premise is that I have an agent that I communicate with directly, who is my Products Team Lead (PTL), in charge of all products. Each product has its own agent that maintains it or builds it if it doesn't exist yet. PTL oversees all of the agents. I never talk directly to any agent other than PTL. I literally have an AI employee that manages about 12 or 13 AI employees. Everything that the agents do to the code gets documented in a journal. Each agent manages a product and a product can consist of one or more projects. Each project has its own Git repo, and each project has its own journal and "sitemap" that gets maintained by the managing agent. Every week PTL collects all of the journals, specs, documents, images, etc., and compiles them into a knowledge base called the knowledge stack. The knowledge stack is such that I can have a conversation with it as the source of truth for my business. BizAgent is not only my products team but it is also the memory for my business. Any kind of news or press release or anything that I give it goes into the knowledge stack. Anything that the agents do to the code goes into the knowledge stack. Every time we spec an idea or project or product, it goes into the knowledge stack. The knowledge stack becomes the memory for the business. **Features** * BizAgent uses the LLM provider's API interface and works with any OpenAI-compatible provider, including Ollama and llama.cpp. When you first run the installer, it asks you for your preferred LLM provider and an API key to access that provider. It then installs all the software and services and gives you a URL to load into your browser. Once you load that URL in your browser, it begins the onboarding process. Using your own LLM provider, it chats with you and asks you more questions about who you are, what your business is, what kind products you want to create, which products are related to other products so the agents can collaborate directly without going through PTL, etc.? * It uses text-to-speech by default for responses so that anything that PTL says, it speaks in a human voice. Combine that with something like Wispr Flow and you could have regular two-way voice conversation with your agent while it delegates work to other agents. * Any Markdown document, PlantUML script, image, or text file created by any agent within any project gets mapped to what's called the library. The library enables you to search and view any file or diagram created within any project. It renders UML diagrams from script and is just a very convenient way to find and read documents. * Because the BizAgent interface is a web interface, the BizAgent doesn't have to reside on your local computer. It can reside on its own dedicated machine or on a VPS. For a business like mine what makes the most sense is to put a dedicated machine on your local network. Put that machine on a VPN so that you can access your BizAgent from outside your office via VPN. * BizAgent and all of the tech that it uses are open source and will always be open source. That's not to say that there won't be a commercial product based on BizAgent but BizAgent as it is now is open source and uses only open source software. I just wanted to let people know what I've got and if it's something that you can use, it's available on GitHub. I personally use it only with Venice AI and local models when I have GPUs available that are not training AI. I use my own GPUs and I run local models. For example I run a Qwen3 27B locally and it works wonderfully with all my agents. I haven't used ChatGPT, Claude, or Grok in months mainly because a lot of the work that I'm doing is confidential and proprietary. I've been burned by those guys in the past when I start working on something that nobody's done before. Then all of a sudden a couple of weeks later, one of them announces this cool new thing that they've come up with (that sounds exactly like the thing that I've been working on while using their LLM provider to build). And Claude was the worst because I was working on some potentially disruptive AI tech using Claude Code and the Claude API. After a while it started sabotaging my project so I just had to stop using it completely. Anyway since then, I haven't used it but now that I've got BizAgent and I've been using Venice and local models, I almost never have reason to go back to the main ones.
document fetched 2026-10-26 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xesttk/i_built_my_own_agent_framework_and_its_awesome