Subscribe
Share
Share
Embed
Hermes Mixture of Agents (MOA): Combine Claude + GPT with an Aggregator to Beat Frontier Models
The script explains Hermes’ Mixture of Agents system, which lets you combine multiple models (e.g., Claude Opus 4.8 and GPT-5.5) into a panel and choose an “aggregator” model to fuse their outputs, treating the mix as one virtual model. The presenter demos results like generating a Windows-style OS and building games, showing side-by-side comparisons where the mixture outperforms a single model, and notes they tested 42 builds viewable on Goldie Bench. It describes how MOA 2.0 works under the hood (private analysis by models, then an aggregator writes the final answer and runs tools), switching mixes with /MOA, and claims of benchmark gains (8% over Opus 4.8, 11% over GPT-5.5). Downsides include slower runtime, API reliance, and technical setup, which they simplify via their Agent OS dashboard, also featuring Fusion and Sakana-style panels, automations, memory, and access via their paid community with tutorials and coaching.
00:00 Mixture of Agents Intro
00:33 How the Panel Works
01:27 Bench Tests and Demos
02:13 Side by Side Game Results
03:23 Tradeoffs and Limitations
04:02 Making MOA Easy to Use
04:45 MOA 2.0 Explained
05:28 Fusion and the Bigger Trend
06:34 Stop Chasing Models
07:10 Three Systems and Why They Win
07:37 How to Use It Today
08:19 Agent OS and Community Pitch
09:50 Wrap Up and Goodbye
The script explains Hermes’ Mixture of Agents system, which lets you combine multiple models (e.g., Claude Opus 4.8 and GPT-5.5) into a panel and choose an “aggregator” model to fuse their outputs, treating the mix as one virtual model. The presenter demos results like generating a Windows-style OS and building games, showing side-by-side comparisons where the mixture outperforms a single model, and notes they tested 42 builds viewable on Goldie Bench. It describes how MOA 2.0 works under the hood (private analysis by models, then an aggregator writes the final answer and runs tools), switching mixes with /MOA, and claims of benchmark gains (8% over Opus 4.8, 11% over GPT-5.5). Downsides include slower runtime, API reliance, and technical setup, which they simplify via their Agent OS dashboard, also featuring Fusion and Sakana-style panels, automations, memory, and access via their paid community with tutorials and coaching.
00:00 Mixture of Agents Intro
00:33 How the Panel Works
01:27 Bench Tests and Demos
02:13 Side by Side Game Results
03:23 Tradeoffs and Limitations
04:02 Making MOA Easy to Use
04:45 MOA 2.0 Explained
05:28 Fusion and the Bigger Trend
06:34 Stop Chasing Models
07:10 Three Systems and Why They Win
07:37 How to Use It Today
08:19 Agent OS and Community Pitch
09:50 Wrap Up and Goodbye
Creators and Guests
What is AI News Today | Julian Goldie Podcast?
Latest Podcast