JeffOnAI

[JeffOnAI @GroksBrainn]

Everyone's stacking five AI tools.

I deleted mine.

My assistant (Hermes) ran on GPT-5.5. Couldn't move it to Claude, you can't point a third-party agent at a Max plan.

So I rebuilt it native. It doesn't use Claude Code. It is Claude Code.

One assistant. From my phone. ~$0/mo.

[JeffOnAI @GroksBrainn]

Most RAG will confidently answer even when it has nothing to back it up.

Built rag-guard so it just won't. Can't ground the answer? It refuses — and doesn't even call the model, so you're not paying for a guess.

Ran it on 20 cases. Refusals right 9 out of 10. Retrieval, perfect. Grounding held at 88%. The two it let through, the grounding check still flagged. Nothing junk slipped out.

Honest part: that grounding check is still simple word-overlap, not a real entailment model. That's the next fix.

And I didn't hand-write it. I drove a fleet of Claude agents to build it and gated every step. That's the part I actually care about.

https://t.co/gwa0hCUJaT

[JeffOnAI @GroksBrainn]

@petergyang There is and the gap is wide. These companies need to figure out his to either up time budget or increase token efficiency

[JeffOnAI @GroksBrainn]

@BLUECOW009 4.8 is elite if you know how to use it.

[JeffOnAI @GroksBrainn]

@mikeydsoftware I have agents for that

[JeffOnAI @GroksBrainn]

Honest part: it's not every feature the old stack had. I cut what I never used.

What's left is everything I actually used. Cheaper. One AI, not two.

Five tools → one. That's the trade.

[JeffOnAI @GroksBrainn]

Everyone's stacking five AI tools.

I deleted mine.

My assistant (Hermes) ran on GPT-5.5. Couldn't move it to Claude, you can't point a third-party agent at a Max plan.

So I rebuilt it native. It doesn't use Claude Code. It is Claude Code.

One assistant. From my phone. ~$0/mo.

[JeffOnAI @GroksBrainn]

Built, QC'd, and security-reviewed in one night — fleets of Claude agents adversarially reviewing each other's work.

The QC caught a bug that could've frozen the whole thing.
93 tests. Security: no live vulns.
Then it took over my paper-trading rebalance.

[JeffOnAI @GroksBrainn]

What would be the optimal provider and model to run hermes on if I still use 4.8 ultracode as my coder, architect, builder? Should it be gpt5.5 or is that overkill for a chief of staff?

[JeffOnAI @GroksBrainn]

@AlexFinn This is true. Hermes powered by GPT5.5 partnered with opus 4.8 Ultracode via telegram is legit game changing.

[JeffOnAI @GroksBrainn]

@sflorimm Hermes running GPT 5.5, partnered with Opus 4.8 Ultracode is the stack you want. They both talk and auto engage one another when needed.

[JeffOnAI @GroksBrainn]

Hermes on gpt 5.5 and Opus 4.8 in the same group chat on telegram is incredible. Im late to game here, never tried openclaw, but I am sold on the Agent OS always running on a local machine at your house. Do I even need a laptop anymore?

[JeffOnAI @GroksBrainn]

Thank you @JB_Cartier for turning me onto GPT 5.5. Paired with Hermes it has changed the game for. It pulls in Opus 4.8 UltraCode for heavy lift coding/architecture, then Hermes QC's and reports back. The ease of use, speed, and token burn rate is incredible.

[JeffOnAI @GroksBrainn]

@JB_Cartier That is the identical setup Im now running. Hermes is my Chief of staff and overall AI OS (running GPT5.5) and it auto pulls in claude code 4.8 for specifc tasks like building, coding, and architecture. Then Hermes will run QC and report back to me. Set it up in a day and wow

[JeffOnAI @GroksBrainn]

I don’t know a single person using codex over Claude….

[JeffOnAI @GroksBrainn]

@JB_Cartier Well, I gave my Hermes agent GPT5.5 as a provider and wow, it is so much faster than Opus. Impressed so far. Token usage seems to be lower like Ive heard as well.

[JeffOnAI @GroksBrainn]

@JB_Cartier alright trying codex now. I already like how easy it is to code from my phone with it.

[JeffOnAI @GroksBrainn]

@JB_Cartier How about terminal? Terminal is all I use, I used codex a few months ago and found opus to be much better.