Learn how to build your own AI agent harness in Rust with a ReAct loop, shell tool execution, a CLI, and local JSONL chat storage. https://t.co/NS7yovMBhS
Just published a new post about AI Agent harness: https://t.co/ET3VT81oCY
In this post I'm explaining how I built AI Coding Agent in Rust and what core components are here.
Last few days I was super busy so didn’t write any code for EliCode. But the prototype is almost done, need to polish and prepare a blog post about it. Hopefully will do it tomorrow
@Heliocene Yeah I tried to build TUI and GUI in Rust as well and realised that it’s very hard. So for now I decided to focus on the core AI Agent with CLI without any *UI due to complexity of building UI.
I decided to return to GUI when will be happy with the core part of an agent
I plan to write next blog posts:
1. How to develop own AI Harness (AI Coding Agent).
2. How to make own AI Harness secure and don't allow to hack my machine.
To write these blog posts I'm developing EliCode - my own AI Coding Agent in Rust.
I don't commit to build GUI for AI Agent right now. It's just a dream to have a high quality GUI for coding agent. Will finish my blog posts and next see what can I do next.
#rust#aiagent#aiharness#buildinpublic
Building custom AI Agent allows me to fully influence the code base and see how it will work.
But I also have a dream around it. I chose Rust because I want to make a good GUI for AI Agent at the target state.
I'm currently writing a new blog post for https://t.co/V2hsXJY3Zz about building own Coding AI Agent in Rust. It will be a simple clone of Claude Code but it's very interesting to build it by hands and understand every part.
But core part of Claude Code clone - AI Agent will be untouched due to limited time which I can use to build it. That's why I changed my plan and focused on the core part of an application - AI Agent with the no UI at all.
All executions work from CLI and print output to stdout.
For the Claude Code clone my original plan was to build extremely fast GUI AI Agent in Rust because Claude Code UI was too slow, but after starting building it with iced GUI library I realised that I need to spend too much time with GUI instead of AI Agents.
Today my Claude Code clone in Rust built the Telegram bot. It's first success execution of it and it works!
For now I support only BashCommand execution so AI Agent can execute any command. As result it produces Markdown file with list of messages and requests to LLM.
As result this AI Agent can be hacked with prompt injection and do with my computer whatever prompt said.
And this is my target actually. I want to simulate prompt injections in my AI Agent and play with AI Agent safety. I'm very excited about it.