Agentic AI and You: Give Your Agent a Brain

A.I Emphasis

Early on, I tested a bunch of smaller models to see which ones could handle agent work. I pointed one at a simple job, installing a tool I use to build knowledge pages for my agents, and its answer, no joke, was “You have a brain and I don’t.” I laughed, then laughed again.

The harness had given that model everything it needed to do the job, and the model just didn’t understand what I was asking. In the agent = model + harness formula from my first article, the model is the brain that reads your request, plans and decides what to do next, and it matters.

Frontier models and open models

Models come in two big families. Frontier models are the most capable ones available; they run in the cloud, and you pay per use. Open models are ones you can download and run on your own hardware. At the very top end, frontier still wins.

Think of it like the 100-meter dash. The frontier model has the best trainer in the world, the best nutrition, the best of everything money can buy, and it runs a 9.0. The open model is just naturally talented, or maybe distilled talent, since many open models learn from bigger models’ answers. A year or so ago it ran a 9.3, then a 9.2, and now it’s running a 9.1.

Coding shows the gap closing most clearly. The best coding models have lived on the frontier platforms, but open coding models have come out that are very, very close. That’s true as of fall 2026, and at this pace, check back in six months.

Four specs worth knowing

You and I aren’t data scientists, so here’s just enough to make a good decision. Each of these turns into a hardware requirement in Post 6.

  • Size. Models are measured in parameters, the numbers a model learns in training, usually in the billions. Bigger models are usually smarter but hungrier for memory.
  • Quantization. Think of it as compression. It shrinks a model so it fits in less memory, at a small cost in quality.
  • Reasoning. Reasoning models think longer before they answer, which gets you better answers but uses more tokens and more time.
  • Context window. How much the model can hold in mind at once. Agents fill it fast, because every loop adds more.

Route each step to the right model

Remember the game-show lifeline where you could phone a friend? Now imagine you get your whole speed dial instead, with the smartest person on every topic, and they always pick up. That’s model routing. Instead of one big model doing everything, your agent sends each step to the model that’s best at it: small models for the routine steps, big models only when you need them. Your agent gets faster, and your bill gets smaller.

I’ve had great results with small, task-specific models like NVIDIA Nemotron Nano and NVIDIA Nemotron 3.5 Lightning, an open model built for the routine work of long-running agents, like calling tools and summarizing results. NVIDIA NeMo Switchyard, an open source routing library, makes the speed dial work by picking the right model for each step automatically. Routine steps are most of the work. When LangChain ran Switchyard on 145 agent tasks, only 7% of the steps needed a frontier model.

Build a shortlist, then run a bake-off

If you need help choosing models to try, start with benchmark sites, which run the same tests on many models and rank the results. Four are worth knowing:

  • Artificial Analysis compares models on quality, price and speed using its own independent tests. Start here to build a shortlist that fits your budget.
  • Arena (formerly LMArena) has real people compare answers from two anonymous models and vote for the better one. It shows which models people prefer for everyday work like writing and answering questions.
  • Berkeley Function Calling Leaderboard tests how accurately a model uses tools like web search and file access. Agents do their work through tools, so for agent work, this is the list that matters most.
  • SWE-bench tests whether a model can fix real bugs in open source software. Check it if your agent will write or fix code.

Benchmarks only get you a shortlist. The real test is your task, so give the agent you built last time the same job on two or three models from that list. You can do it two ways:

In the cloud. One OpenRouter account gets you models from many providers, and in Hermes you switch between them with hermes model. Try one frontier model and one or two smaller open ones.

On your own computer. Hermes can use LM Studio as its model provider, so the model you loaded in my first article can power your agent for free. Small models may stumble on multi-step work. Watch them when testing to understand if they fail and exactly where.

Score each run on four things: did it finish, was it right, how much feedback did it need, and what did it cost. Keep the cheapest model that gets it right, and save the frontier model for the steps that need it.

Show your work: what I tested and what I run

  • Tested: A lot of smaller models. One of them told me I had a brain and it didn’t.
  • Build: Frontier models through OpenRouter and my Claude plan. Designing an agent and writing its skills is the hard thinking, and it’s worth paying for.
  • Run: Once a skill is built, it runs every day on DeepSeek V4 Flash, an open model, on my Dell Pro Precision with GB10, an AI workstation that sits on my desk. Right now, it’s the ticket.
  • Caveat: I’m always testing, so ask me again next quarter.

When this isn’t the right answer

Sometimes only a frontier model will do. If your bake-off shows that only the frontier model gets the job right, or the task is the hardest kind of coding, reasoning or research, pay for the best. Don’t force an open model into a race it can’t win yet.

Now your agent has a brain. Next time, we’ll build the body: hands to do the work, memory so it doesn’t start from scratch every time, and a routine so it runs while you sleep.

Dell reported this
Source: www.dell.com
Source link

Leave a Reply

Your email address will not be published. Required fields are marked *

5 × 2 =