Update - March 24, 2026: I've been working with GLM-5 and the just-released Minimax M2.7 for a bit now and they're giving me great results — even against oh-my-openagent's recommended config. Given that, OpenCode Go would be plenty of subscription for most people. But even if you were paying per token, I have multiple significant projects being solely built by agents and have spent about $50 this month.

The Problem

Everyone wants you to know how to do agentic coding now. But time and space to learn it at work can be scarce, and tokens are expensive.

These are basically exercise starting places, not end to end how to guides. You might need to go on your own adventure a little bit. These also assume you're a software engineer and may already know some things like how to navigate the command line, how to write an API client or otherwise find and use an API client, etc.

If you want the broader version of how I'm thinking about AI right now, see It took me a year to write about AI and all I got was this stupid blog post....

Use OpenCode Go / OpenCode Zen with Oh My Openagent

This is one of the most affordable ways to get decent performance out of your models with very little setup and customization that I know.

Opencode is an agentic coding tool like Claude Code, Codex, or Gemini CLI.

They offer routing to many popular agents through a service called Zen and they offer cheaper alternative models under a smaller subscription called Go (we're talking $10 a month as of this writing). This isn't an advertisement, I have no association with them, it's just the only dirt cheap way I have found so far to get something decent. If you can afford $25-50 of tokens at a time and learn with small projects monthly or so, do it. Having access to the frontier models can open your eyes to what "pairing" with a really good Genie, as Kent Beck calls them, can get you.

Once you have Opencode installed and you've setup at least a Go subscription, check out this project: https://github.com/code-yeongyu/oh-my-openagent

To install it, you just prompt a model you have available on Opencode Go like the Kimi 2.5 model or GLM-5. When you step through the prompts it'll suggest setting up a bunch of accounts, you can say no and instead tell it you have access to OpenCode go and have it setup something that makes sense.

Now, oh-my-openagent is going to setup a LOT of magic for you... but it's magic that works well. Opencode includes a lot of customized agents with lengthy prompts to guide their workflows so they check and balance each other to develop thorough plans, collaborate in a swarm... it's very powerful. You will probably use it in a way that gets you bad results at first. That's ok. Try to find a nice rhythm of asking it questions, steering it towards a good plan that feels well designed and verifying quality & functionality seems reasonable, then ask it to execute the plan and assess what it did.

As you start to get better results, dig into oh-my-openagent's prompts, documentation, and start to wrap your head around how it works and its design principles.

Build anything you've wanted to build and haven't... A work side project from 3 jobs ago that never manifested, an update to your own blog, something to keep your family's calendars synced (I did that). See where you can land.

A note on frontier models

Frontier models are models like the latest Claude Opus, the latest GPT, etc.

You will get better results with less effort from those models. It just is. If this concerns you at all let me reassure you of a couple things: 1. As best I can tell from my own biased experience, these cheaper models are about where Claude was between 6 months and a year ago. People were still doing agentic code then, with a bit more hassle. 2. Every time an agent veers off in a wrong direction is an opportunity to learn about what context can set agents up super well. Most of keeping the agent on track is about managing context in someway whether it's pointing the agent at clearer context, context more optimized for it (see vector databases), or managing it's context window (how much stuff it can keep in it's head at once without getting confused and overwhelmed... like us). 3. You can always selectively spend 25 bucks at a time on an afternoon of planning or other complex tasks with Claude Opus or another big frontier model to feel what the difference is and get a sense of how you might be able to adjust your approach on the job w/ access to more expensive models.*

Reflection Questions

Ask yourself these questions as you practice: 1. Where did the risk move from when I was handwriting the code? 2. What options do I have that I already know and are familiar that might help me manage that risk? 3. Does this open up new options for managing risk, or even new unmanaged risks? Is there anything to learn from another way of building software about how to manage this new risk? 4. How can you keep this context in your head now? Which context is important to keep? How can you improve the communication between you and the agent about what it did? Can you use the agent to learn about what it's doing as it works? 5. What does it take to land on something that is easy to maintain. Something where you feel confident making changes, can verify that things work, can deploy it with a normal level of security concerns? 6. What does it take to land on something that brings other people value. That's why we do this after all... You can have other reasons for having your job, affording your life, taking care of your family... but the reason people need us to do our job is to make their lives slightly better. I don't mean investors. I don't mean executives. I mean the folks who use the software we write.

Run Your Own Models

For those who are adventurous, have extra time on their hands, who learn better with more ambitious projects, or just need to understand how everything works under the hood to really feel like you "get it", learn to run your own models. This is where I started! It was maybe slightly less painful in that I have a devops heavy background (remember devops?).

But the technology is changing fast, there's a thousand new tools every day, and figuring out how to run a decent model on the hardware you can have, not for the faint of heart.

That said, if you have already spent a ton of money on a machine you've heard is capable of running decent models or you're a huge DIY nerd who will be motivated by going this hard route, here's a couple early things to try:

Ollama

https://ollama.com/

Why? It's available, it's pretty well documented, it's been around for a bit and feels more mature than a lot of other tools. There's a nice UI if you like but you can also do a lot at the command line.

Which Models?

This part is hard, might slow you down for a minute to learn about models and quantization for a while.

What's quantization? This Digital Ocean article can walk you through what it is but here's the gist from the article:

Quantization is a technique that reduces the precision of model parameters, thereby decreasing the number of bits needed to store each parameter. For instance, consider a parameter with a 32-bit precision value of 7.892345678. This value can be approximated as the integer 8 using 8-bit precision. This process significantly reduces the model size, enabling faster execution on devices with limited memory.

TL;DR: Quantization lets you run a model with less RAM.

Take a look at this blog for some examples of what models you might be able to run at what quantization with what kind of hardware: https://emelia.io/hub/run-ai-locally-hardware-guide-2026.

I'd recommend ones that try to compete with big frontier models like Anthropic's Claude models. Some examples to keep an eye out for are the latest qwen model and devstral.

https://huggingface.co/ is a great source for finding quantized models and downloading them. Think of it as the github for ML & LLMs.

Now What?

I always start with querying the model in the simplest way I can. At the command line or in the UI (read the Ollama docs).

Once you've confirmed the damn thing works, figure out how to hit its API. Then review documentation for a library like langchain or Pydantic AI. Research similar clients for a language you're familiar with but langchain has a typescript client for many of us who are familiar with typescript / javascript.

Interacting with the API more directly can teach you a lot about how LLM-driven AI agents work and that can really help you when you use a tool like Claude Code or similar at work.

Once you've got that going you're going to have to figure out how to point a tool like (Opencode)[https://opencode.ai/] to it! It should be pretty straight forward. If in doubt, Opencode is automatically set to talk to a free model and it might be able to help you figure out how to configure it.

Just Do It

(Can you tell I briefly worked at Nike?)

Seriously though...

Just Do It - Shia LaBeouf

It's time to learn. I know there's so much chatter right now about AI making people hyper productive or shipping code like they've never shipped.

Maybe I'll eat my own words one day but, in my best jobs with the most clarity of purpose and clear runway, I shipped more compared to what I can ship in a typical enterprise navigating project plans and alignment across teams and orgs of people. Straight into users hands. And on my own homelab projects, I'm currently spinning up swarms of agents and pointing agents at tickets I wrote and figuring out what I need to do to build a safe system around them so I don't need to worry about the exact code they write.

I've found agentic coding to be more about moving risk, spontaneous value from unpredictability, shifting my thinking in creative ways to think about the system design, and spending more time thinking about the purpose of what I'm doing than the minutia of what I'm doing.

And I love the minutia of what I'm doing. I'm a huge language nerd and was an early adopter of most programming languages over the last 15 years from Elixir, Gleam, Clojure, Elm, my friends' toy languages... I love crafting code. I love doing it with people.

This is a new tool to figure out. None of us really understand yet all the ways it's changing the game but I can tell you there are lessons in here worth learning.

Footnote