I'm mildly surprised that more people aren't using Nemo models for this reason. We've moved most of our processing to a combination of Nemo Ultra and Super, with some support for multi-model-specific tasks on Omni. The setup is working REALLY well for us, and I'm comfortable with the more measured pace of improvements. We work with many long-context problems, and the ecosystem is great.
There were a number of use cases where we needed to use Gemini (audio modality), and Ultra has been a VERY cost-effective alternative once we got through the nuances.
I'll pipe in here as someone working on an agentic harness project using mastra as the harness.
Nemotron3-super is, without question, my favorite model now for my agentic use cases. The closest model I would compare it to, in vibe and feel, is the Qwen family but this thing has an ability to hold attention through complicated (often noisy) agentic environments and I'm sometimes finding myself checking that i'm not on a frontier model.
I now just rent a Dual B6000 on a full-time basis for myself for all my stuff; this is the backbone of my "base" agentic workload, and I only step up to stronger models in rare situations in my pipelines.
The biggest thing with this model, I've found, is just making sure my environment is set up correctly; the temps and templates need to be exactly right. I've had hit-or-miss with OpenRouter. But running this model on a B6000 from Vast with a native NVFP4 model weight from Nvidia, it's really good. (2500 peak tokens/sec on that setup) batching. about 100/s 1-request, 250k context. :)
I can run on a single B6000 up to about 120k context reliably but really this thing SCREAMS on a dual-b6000. (I'm close to just ordering a couple for myself it's working so well).
Good luck .. (Sometimes I feel like I'm the crazy guy in the woods loving this model so much, I'm not sure why more people aren't jumping on it..)
I think the entire premise that the prompting is the surface area for optimizing the application is fundamentally the wrong framing, in the same way that in 1998 better cpam will save CGI. It's solving the wrong problems now, and the limitations in context and model intelligence require a tool like Dspy.
The only thing I'd grab dspy for at this point is to automate the edges of the agentic pipeline that could be improved with RL patterns. But if that is true, you're really shorting yourself by giving your domain DSPY. You should be building your own RL learning loops.
My experience: If you find yourself reaching for a tool like Dspy, you might be sitting on a scenario where reinforcement learning approaches would help even further up the stack than your prompts, and you're probably missing where the real optimization win is. (Think bigger)
You just articulated why I struggle to personally connect with Gemini. It feels so unrelatable and exhausting to read its output. I prefer to read Opus/Deepseek/GLM over Gemini, Qwen and the open source GPT models. Maybe it is RLHF that is creating my distaste from using it. (I pay for Gemini; I should be using it more... but the outputs just bug me and feel more work to get actionable insight.)
It's getting a lot easier to do this using sub-agents with tools in Claude. I have a fleet of Mastra agents (TypeScript). I use those agents inside my project as CLI tools to do repetitive tasks that gobble tokens such as scanning code, web search, library search, and even SourceGraph traversal.
Overall, it's allowed me to maintain more consistent workflows as I'm less dependent on Opus. Now that Mastra has introduced the concept of Workspaces, which allow for more agentic development, this approach has become even more powerful.
I love where you're going with this. In my experience it's not about a different persona, it's about constantly considering context that triggers, different activations enhance a different outcome. You can achieve the same thing, of course by switching to an agent with a separate persona, but you can also get it simply by injecting new context, or forcing the agent to consider something new. I feel like this concept gets cargo-culted a little bit.
I personally have moved to a pattern where i use mastra-agents in my project to achieve this. I've slowly shifted the bulk of the code research and web research to my internal tools (built with small typescript agents).. I can now really easily bounce between different tools such as claude, codex, opencode and my coding tools are spending more time orchestrating work than doing the work themselves.
The skepticism is understandable given the trajectory of GPTs and custom instructions, but there's a meaningful technical difference here: the Apps SDK is built on the Model Context Protocol (MCP), which is an open specification rather than a proprietary format.
MCP standardizes how LLM clients connect to external tools—defining wire formats, authentication flows, and metadata schemas. This means apps you build aren't inherently ChatGPT-specific; they're MCP servers that could work with any MCP-compatible client. The protocol is transport-agnostic and self-describing, with official Python and TypeScript SDKs already available.
That said, the "build our platform" criticism isn't entirely off base. While the protocol is open, practical adoption still depends heavily on ChatGPT's distribution and whether other LLM providers actually implement MCP clients. The real test will be whether this becomes a genuine cross-platform standard or just another way to contribute to OpenAI's ecosystem.
The technical primitives (tool discovery, structured content return, embedded UI resources) are solid and address real integration problems. Whether it succeeds likely depends more on ecosystem dynamics than technical merit.
There were a number of use cases where we needed to use Gemini (audio modality), and Ultra has been a VERY cost-effective alternative once we got through the nuances.