← Back to context

Comment by sim04ful

14 hours ago

You’ve done an incredible amount of work — I’m definitely going to try out the Zed integration.

I’m curious though, how significant do you think it is for the agent to have semantic access through Tree-sitter?

Also what model have you had the most success with ?

3 comments

sim04ful

Reply
vinhnx  13 hours ago

Thank you for your very kind-words. I love building and agentic coding is my current curiosity.

> I’m curious though, how significant do you think it is for the agent to have semantic access through Tree-sitter?

For this, I'm really not sure, but since the start of building VT Code. I just had this idea to use tree-sitter to assist the agent to have more (or faster/more precise) semantic understanding of the coding, instead of relying them to figure out themself. For me, naively I think this could help agent to have better language-specific and accurately decision about the workspace (context) that they are working. If not having tree-sitter, I think the agent could eventually figure out itself. For this aspect, I should be research more on this topic. In VT Code, I included 6 language: Go, Python, Rust, TypeScript, Swift... ) via rust-binding crates, mostly when you launch the vtcode agent on any workspace, It will show the main languages in the workspace right way.

> Also what model have you had the most success with ?

I'm having mainly limited-budget so I can only use OpenRouter and utilize its vast amount models support. So that I can prototype quickly, for different use-cases. For VT Code agent, I'm using mainly x-ai/grok-code-fast-1, in my experience, it most suit for building VT Code agent it self because of speeds, and versatile in function calling and have good instruction following. I also have good successes with x-ai/grok-4-fast. I have not tried claude-4.5-sonnet and gpt-5/gpt-5-codex though. I really love to run benchmarks for VT Code to see how it perform in real world coding task, I'm aiming for Aider polygot bench, terminal-bench and swe-bench-lite, it is in my plan for now in my GitHub issues.

For VT Code itself, I instruct it to strictly follow system-prompt, in which I take various inspiration from Anthropic, OpenAI and Devin guide/blogs on how to build coding agent. But, for a model-agnostic agent, the capability to support multi providers and multi models is a challenge. For this I think I need help. I'm fortunately to have support from open-source community suggesting me to use zig, I have had good success with it so far, for implement LLM calls and implement the /model picker.

Overall in my experience building VT, the most important aspect of effective coding agent is context engineering, like all big-lab has research. A good system prompt is also very important, but not context is everything. https://github.com/vinhnx/vtcode/blob/main/prompts/system.md

// Sorry, English is not my main language, so pardon the typo and grammar. Thank you!