Setting aside the AI-slop products advertised on YouTube (personal AI OS, Biz AI OS, etc.) — is there a reason you haven't built a centralized command center that incorporates your CRM, to-dos, ideas, etc.? Something like a 'Hermes agent'? Is it a lack of economic value, or something else?
I basically already have one — it's Linear (a project management/issue-tracking platform, usually used by coding teams), integrated one-for-one with Claude. I use Claude to both track my to-dos and manage them: if I tag something 'agent ready,' it automatically gets delegated to an agent. I have subgroups for different areas too — like a content pipeline I can delegate pieces of to people or to agents running skills. It's really just a run-of-the-mill CRM with a bunch of agentic tools layered on — but honestly, that's all you actually need.
The best way to create agents with a team right now is something like Bugs, where you can vary it with tools like Slack, Claw, Tag, Linear, or Cursor, adding more functions as needed, and you're okay using those models. If you want a framework to collaboratively use agents within teams and want those agents to be able to do whatever you need, check out Buzz Agent. It's a beautiful yellow-associated tool that lets you use virtually any sort of agent and looks similar to Slack. For example, the Bumble New Direction team moved their prototype to Flutter, and the Bumble agent indicates they're doing things right now. You can use it with any agents you want; I've signed up and find it visually beautiful and stimulating, with no major downsides. The problem is already solved: it's just a matter of building system prompts so each persistent agent has its own identity and high quality.
ai agentsproject managementcollaborationbuzz agent
The AI agent I use most is Claude, specifically Claude code (cloud code). It handles most of what I need, even though it can be verbose and sometimes requires scrolling through output. I connect Claude to various platforms via built‑in plugins or connectors, give it access to my email and other tools, and use remote control to have a virtual AI assistant on my phone, computer, etc. I also use Codex and occasionally other tools, but Claude remains my primary assistant.
They’re essentially free to run—about $20 a month plus any token usage. If you want to replicate the setup, just follow one of my courses from start to finish. The agent is just the harness with connectors and context for your environment. Platforms like Grockbot abstract away credential management, making it easy to connect everything and unlock more value.
The vast majority of those elaborate AI‑agent flows you see online are just for views—they don’t actually make money because the agents aren’t reliable enough to do real work. Businesses don’t want a model that can answer in a million different ways; they want a highly constrained output—maybe just five or six possible answers—so they can control most of the content. What they really want is for you to scrape a resource, run it through an LLM API to get a few rephrased versions (like icebreakers or paraphrased job titles), and then slot those into a templated email or document. The LLM fills in only the small, variable pieces while the rest stays fixed, giving them control and reducing unpredictability. In short, treat LLMs as APIs rather than autonomous agents; that’s where the money is right now.