I've tested it extensively last year but I got into repeated fights with Goose's tooling logic so I've stopped trying for now. I was planning to compare some frameworks when the announced `jan-v3-coder` model launches.

I've been curious about Goose after all the hype it's gotten from the presidio bitcoin podcast. I haven't tried it yet, but my experience with llama 3.3 and Qwen3 hasn't been great. There's still a fair rate of hallucination in both, which can become tedious to correct.  I find llama3.3 can also be easily distracted by details unrelated to the task at hand. I guess that's fair as it can happen to me too.

The fact that agentic coding works at all and that you can run open source models to do it is kind of amazing though.

endothermicdev

I was curious if Block's Goose agent, paired with Ollama and the Qwen3-coder model, could really replace Claude Code. Here's how it worked.