I am finally happy with my agentic dev setup without using costly MaaS. This website was created in 3 hours running a model on DGX Spark.
- 15 minutes of human‑agent conversation to get the spec right.
- 2 h of unsupervised agent work.
- 45 min of human testing and agent fixes.
What did I use
The tech stack I can really recommend for agentic engineering is:
- Opencode as an agent
- Qwen 3.6 LLM in a 35‑billion‑parameter variant
- docker‑sbx for a sandbox
Agent
I have been using GitHub Copilot for quite a while. As an experienced developer I was originally very skeptical towards “vibe coding”. It is alright for smaller, well‑defined development tasks, but when you move to spec‑driven development and larger tasks, the produced output becomes overly complicated and difficult to inspect—regardless of the selected model. Development with open‑weight models fails quickly.
Opencode seems to solve all of the issues above. It works really well with OW models; the produced code is lean and allows me to keep track of my context‑window size.
Model
Qwen 3.6 gets all the hype, and this is deserved. The model runs very quickly in Ollama on DGX Spark. The context window is 262 144 tokens, which is quite sufficient for a long development session. And most importantly:
- it produces good‑quality code;
- it actually finds and resolves problems as they arise.
I also have a good experience with Google’s Gemma 4 in a 31‑billion‑parameter variant, but it is much slower than the Qwen family.
Sandbox
We’ve all read the stories about agents deleting important files or bricking the OS they run on, and that made me look for a sandboxing solution. It was obvious that a VM was needed, since you need access to many tools and OS features when you create software.
I was ready to start testing Vagrant, which, by the way, is still maintained by HashiCorp. Then I discovered docker‑sbx. The solution is pretty neat and lightweight. Docker provides slim images for most popular agents, and deployment is just one command.
Summary
I hope this was useful, especially if you are worried about your token bills.
Contact us to discuss your GenAI strategy.