Agentic AI in 2026: Hermes, NVIDIA Nemotron and the open-model stack

The interesting frontier in 2026 is not a bigger chatbot. It is the agent: a model that plans, calls tools, reads the results, and loops until a task is done. Chat answers a question; an agent changes the state of a system. And the moment a model can act, reliability stops being a nicety and becomes the entire product.

Two open-weight families have become the backbone of self-hosted agents — Nous Research’s Hermes and NVIDIA’s Nemotron. Here is why, and what the stack around them looks like.

What “agentic” really requires

A capable conversational model is not automatically a capable agent. Agents live or die on a narrower set of skills:

• Function calling — emitting well-formed, schema-valid tool calls, every time, not most of the time.
• Structured output — reliable JSON the surrounding code can parse without heuristics.
• Steerability — actually obeying the system prompt that defines the agent’s role and limits.
• Reasoning under a loop — planning multi-step work and recovering when a tool returns an error.

A model that is 2% better at trivia but 20% worse at emitting valid tool calls is a worse agent. That single fact explains most of the open-model landscape.

Hermes: the open, steerable workhorse

Hermes, from Nous Research, is a family of fine-tunes built on open bases such as Llama. Its reputation is built on three things agent builders care about: strong function calling, dependable structured/JSON output, and a deliberately neutral, steerable alignment — it follows your system prompt rather than a vendor’s hard-coded persona.

For self-hosted agents that need to slot into your own guardrails and behave predictably, that steerability is the whole appeal. You define the role; the model adheres to it. It is the same instinct behind running models locally that I explored earlier — control over the behaviour, not just the weights.

NVIDIA Nemotron: reasoning tuned for tools

Nemotron is NVIDIA’s open-model line, and the Llama Nemotron reasoning models (in Nano, Super and Ultra sizes) are explicitly tuned for agentic work: multi-step reasoning and tool calling, with reasoning that can be toggled on for hard tasks and off for cheap ones. They ship with open weights and are packaged as NIM microservices for deployment behind a standard API.

The Nemotron family also earns its keep in an unglamorous place: synthetic data generation. Larger reward and instruct variants are widely used to generate and grade training data for other models — including agents you might fine-tune yourself.

The stack around the model

The model is maybe a third of a working agent. The rest is engineering:

• Tool schemas — tight, well-described function definitions. Vague schemas produce vague calls.
• Orchestration — a plan/act/observe loop (ReAct-style), with step limits and timeouts so a confused agent fails fast instead of burning tokens forever.
• Guardrails — validate every tool call before it executes; never let the model reach production systems unchecked.
• Evals — a golden set of tasks you re-run on every model or prompt change. Without evals you are not engineering, you are guessing.
• Observability — trace every step. When an agent misbehaves at 2 a.m., the trace is the only thing that saves you.

Open vs API: how to choose

Frontier API models are still the easiest way to get a capable agent working today, and for many teams that is the right call. Reach for the open stack — Hermes, Nemotron — when you need data residency, predictable cost at high volume, deep customisation through fine-tuning, or simply the guarantee that the model will not change under you next quarter. Often the pragmatic answer is a mix: a frontier model for the hard reasoning, an open model for the high-frequency, well-scoped calls.

---

Agentic AI rewards boring engineering discipline far more than model hype: tight tools, a strict loop, real guardrails, and evals you trust. Pick a model that is genuinely good at tool calling — Hermes and Nemotron both are — and spend your energy on the harness around it. If you are wiring an agent to real business data, start where the data lives: the Databricks lakehouse and Fabric with Copilot Studio.