Great breakdown of the tool layer. I have been comparing agent harnesses myself — Codex, Claude Code, Pi, OpenClaw, Hermes, Reasonix, Kilo Code, and others. If your readers want to stay on top of which new harnesses are shipping and how they compare, we cover this daily in the Agent Builders newsletter: aiagentbuilders.substack.com
I have to say that the Anthropic's models harness is very high compare to Deepseek models. I use Kilo Code and Kilo CLI to achieve the same goals but with a great difference in tokens consuming. So, I highly recommend the stack Kilo/Deepseek.
I have a question regarding the practical application of this approach. Do you see a possibility of building an AI agent that can be trained on a company’s specific coding standards, proprietary functions, internal packages, and development frameworks, so that it can assist developers in generating code from prompts based on the existing codebase?
For example, we are a product company that develops R packages for specific modeling products. We follow a well-defined framework consisting of a collection of internal packages, configuration standards, custom model-production scripts, reusable functions, and established practices that help maintain consistency across different model implementations.
Given the tools and approaches discussed here, would it be possible to provide this framework and codebase as context to a custom agent, enabling it to generate code that adheres to our company standards and best practices? The goal would be to improve developer productivity, accelerate development, and ensure consistency with our existing architecture and coding conventions.
As always, an incredible post! Thanks a ton Hadley! Just as an idea for future posts… it would be incredible to see how skills could be incorporated into ellmer workflows!
Thanks for teasing out how a coding agent work, Hadley. It helps me to write my own and harness other coding agents like gemini-cli, claude-code, etc
Super nice write up. Thank you!
You've probably seen it, but if not you might like this similar-in-spirit post from Amp: https://ampcode.com/notes/how-to-build-an-agent
Nice post! Thanks for sharing.
Great read. We featured it on episode 1 of our podcast. Thanks for making content like this! https://attentiondeficitpod.substack.com/p/welcome-to-attention-deficit
Great breakdown of the tool layer. I have been comparing agent harnesses myself — Codex, Claude Code, Pi, OpenClaw, Hermes, Reasonix, Kilo Code, and others. If your readers want to stay on top of which new harnesses are shipping and how they compare, we cover this daily in the Agent Builders newsletter: aiagentbuilders.substack.com
I have to say that the Anthropic's models harness is very high compare to Deepseek models. I use Kilo Code and Kilo CLI to achieve the same goals but with a great difference in tokens consuming. So, I highly recommend the stack Kilo/Deepseek.
Hopefully this series will lead to how posit ai is built! Only tried it briefly but quite like it. Nothing else comes close for data work
Amazing 👌 Looking forward to the next part.
I have a question regarding the practical application of this approach. Do you see a possibility of building an AI agent that can be trained on a company’s specific coding standards, proprietary functions, internal packages, and development frameworks, so that it can assist developers in generating code from prompts based on the existing codebase?
For example, we are a product company that develops R packages for specific modeling products. We follow a well-defined framework consisting of a collection of internal packages, configuration standards, custom model-production scripts, reusable functions, and established practices that help maintain consistency across different model implementations.
Given the tools and approaches discussed here, would it be possible to provide this framework and codebase as context to a custom agent, enabling it to generate code that adheres to our company standards and best practices? The goal would be to improve developer productivity, accelerate development, and ensure consistency with our existing architecture and coding conventions.
Yes, definitely, but that's something that you'd do mostly via the prompt/skills
As always, an incredible post! Thanks a ton Hadley! Just as an idea for future posts… it would be incredible to see how skills could be incorporated into ellmer workflows!
Ooh yes, I should talk about skills in the future!