Last month I asked a free AI model running on my own computer to install a writing-voice skill into my agent’s brain.
It failed twice.
The first failure was me: I was in chat mode, not agent mode. (Ask me how long that took to notice...) The second failure was funnier. Instead of installing the skill, the model started roleplaying it. It read my voice guide and began talking like me.
(Flattering. Not useful.)
So I handed the same file to Claude Opus 4.8, a frontier model (the newest, most capable tier of AI), and it installed the skill in one shot. Then it tested itself: 80% confidence, bumped to 95% after a second pass.
That failure taught me more about AI agents for small business than any feature list, because it exposed the pattern that makes agents affordable: build with the expensive model, run on the cheap one.
If your current setup is a ChatGPT tab plus a pile of disconnected subscriptions — juggling tools, no system — what follows is the plain-language version: what an agent actually is, what mine cost to run, and the first build I’d hand a non-technical owner.
What an AI agent actually is
A chatbot answers questions. An agent does jobs: it takes an objective, breaks it into steps, and works through them, opening files, using tools, and checking its own output along the way.
For getting started, two clearly distinct kinds are worth knowing.
Workspace agentslive in an app and work alongside you: chat, scheduled tasks, calendar, email, research. The one I’ve been testing is Odysseus, PewDiePie’s open-source project (60,000 GitHub stars, which is the trust signal I check before installing any random repo).
Build agents live in a folder and make files. OpenCode (163,000 stars) runs in your terminal; Claude Code, the one I teach with, works the same way. Point one at an empty folder, ask for a landing page, and watch the files appear.
The lines blur once you get deeper. As an on-ramp, though, the split is the clearest mental model I’ve found: one type manages your day, the other builds your tools.
Both kinds live inside a harness: the instructions, files, and guardrails you wrap around a model so it does reliable work instead of improvising. If that word is new, my plain-language guide to what an AI harness is covers the whole architecture.
One habit worth copying from day one: I ran both agents in Docker containers, sandboxes that cap what a program can see. My Odysseus container gets 7.5 GB of memory and no idea what computer it’s on. If an agent goes rogue, it goes rogue inside a box.
Build with the $30 model, run on the $4.66 one
Back to that failed skill install, because the fix is the whole strategy.
Odysseus encodes it natively as the “teacher model”: when the cheap local student fails a task, the system escalates to a state-of-the-art teacher that writes a step-by-step skill file, so the cheap model succeeds next time.
That happened in my session without me planning it. Gemma 4 flailed, Claude Opus 4.8 wrote the skill, and now the free model has a recipe it can follow forever.
The per-token math settles it. Per million tokens (the metered units AI usage is billed in), from Venice’s model page:
- Kimi K2.6: $0.85 in, $4.66 out
- Claude Opus: $6 in, $30 out
- GPT 5.3 Codex: $2.19 in, $70.50 out
Opus output costs about 6.4x Kimi’s. But you pay the frontier rate once, at build time... then the cheap model runs that workflow every day after.
When you’re building a tool for your business, use the best model available. Opus costs me $30 per million output tokens, and it hasn’t missed once on these builds.
One caution for anyone budgeting around AI: today’s subscription prices are subsidized. My $200/month Claude Code subscription includes an amount of Opus that can’t be sustainable for Anthropic.
I don’t know what the correction looks like — price hikes, deprecated models, gated access — and I hold that theory loosely. Budgeting at real per-token prices instead of subsidized subscription prices is the hedge either way.
The privacy question the brochures skip
Now the part that matters most for a business: where your data goes.
The big providers store your prompts and may train on them — data is the business model. Paste in customer emails, financials, or a client contract, and you’ve shared them with a company whose product improves by reading them.
Two ways out, and you can use both at once.
Go local. The compressed (quantized) version of Gemma 4 12B needs about 7.4 GB of memory and reasons like models that need 30. 8 GB of RAM might be enough; 16 GB is comfortable; 32 GB lets you step up to Gemma 4 31B. Ollama is the simplest way to run it, and nothing you type ever leaves your machine.
Go private-by-contract.At the time of this build, Venice routed 89 models behind one API key (the catalog keeps growing), each marked private (zero-data-retention agreements with the provider) or anonymized (Venice proxies your prompts like a VPN, so the provider never sees who’s asking). Full transparency: I work with Venice. The architecture is the point, though. Pick a provider whose business model is something other than your data.
Underneath both options sits a belief we hold at the Academy, stated as a belief: own your agents, swap your models. The agent you build (the instructions, skills, and workflows) is yours. The model underneath is a rental, and rentals get swapped.
Your first agent, without touching code
You don’t need to program any of this. When OpenCode built me a landing page and rendered my headshot comically large, the entire fix was one sentence: “the image is too big, it doesn’t fit, fix it.”
You describe, it corrects. That’s the interface now.
The on-ramp I’d hand a business owner, in order:
- Start with a build agent. Watching files appear in a folder makes agents concrete faster than any explanation. My Claude Code guide for non-technical builders starts before beginner: what a terminal is, what it costs, how to set up without panic.
- Give it one real job. Not a demo — a task your business repeats every week. The harness build tutorial walks through wrapping that job in instructions the agent follows the same way every time.
- Then, if the privacy math matters, go local for a weekend. Install Docker Desktop, add Odysseus from its GitHub repo (check the stars before pasting any install command... 60,000 is a good sign), pull Gemma 4 12B through Ollama, and set a frontier model as the teacher for whatever the local model fumbles.
So you can calibrate expectations, my growing pains from one week:
- Ollama showed offline inside Odysseus until I stopped and restarted the Docker container.
- Adding a local endpoint as a custom URL? I haven’t gotten it to work. I use the built-in scan instead.
- In a blind four-model comparison, two of the four models flat-out failed. I don’t know why.
Expect an evening of fiddling rather than a polished onboarding flow, in other words. Every snag on that list was solved (or dodged) with a restart or a plain-English workaround, and none of it required code.
The honest verdict
Worth it? For me, yes, with caveats.
The local model failed exactly where you’d expect a 7.4 GB model to fail: on a multi-step agentic task. It also handled chat, research, and scheduled email drafts without complaint. The frontier model cost money and did the hard 5% perfectly.
What I’m less sure about is whether open-source models close the gap this year. I think they will... but I’ve been wrong about timing before — it took me a year to notice AI could edit video for me.
What I am sure about is the direction. This is what AI automation for small business looks like when you own it: workflows built once with a strong model, run daily on a cheap one, with your data staying on your terms. Stack enough of those and the goal stops being theoretical: the systems carry the workload, and your week opens back up.
And if you’d rather not debug the Docker container alone, that’s the gap the Academy exists to fill: a clear order to learn in, feedback on your real builds, and two community calls a week with fellow builders mid-experiment ($83/month or $497/year, 7-day free trial). Bring your half-broken containers.
Questions people actually ask
- What is an AI agent for a small business?
- An AI agent is software that completes multi-step work on its own: it can read and write files, draft and schedule emails, do research, and build simple tools. A chatbot answers questions; an agent does jobs. For a small business, that means repetitive work — weekly summaries, content repurposing, drafting replies — can run as a system you own instead of a task you perform by hand.
- What's the difference between a chatbot and an AI agent?
- A chatbot waits for a message and replies with text, and you carry the answers between tools yourself. An agent takes an objective and works toward it in steps: opening files, using tools, checking its own output. Two beginner-friendly kinds exist — workspace agents that manage tasks alongside you (chat, calendar, email), and build agents that live in a folder and create files.
- How much does it cost to run an AI agent?
- Model usage is billed per token, the metered units of AI text. At mid-2026 prices on Venice, Kimi K2.6 costs $0.85 per million tokens in and $4.66 out, while Claude Opus runs $6 in and $30 out. The pattern that keeps costs sane: pay frontier rates only while building a workflow, then let a cheap or free local model run it every day after.
- Can I use AI agents without knowing how to code?
- Yes — the interface is plain English. When an agent built me a landing page and rendered my headshot comically large, the entire fix was one sentence: "the image is too big, it doesn't fit, fix it." You describe, it corrects. The real skills are describing what you want clearly and checking the result, not programming.
- Are AI agents private enough for business data?
- They can be, if you choose where the model runs. The big providers store prompts and may train on them — data is their business model. Two ways out: run a local model like Gemma 4 12B, which needs only about 7.4 GB, or use a provider with zero-data-retention agreements. Running agents inside a Docker container also limits what they can see on your machine.
- What computer do I need to run a local AI model?
- Less than you'd think. The compressed (quantized) version of Gemma 4 12B needs about 7.4 GB of memory yet reasons like models needing 30. 8 GB of RAM might be enough, 16 GB is comfortable, and 32 GB lets you step up to Gemma 4 31B. Ollama is the simplest way to download and run it.
