Rishi

Building Kiln - a light weight agent harness tool

[Rishi @RishivikramN]

@Samaytwt Lol, I switched to raycast longtime ago

[Rishi @RishivikramN]

@elponick That’s true, im trying to find a painpoint in dev tools domain but no luck so far. Wbu? Do you have any ideas that you want to work for?

[Rishi @RishivikramN]

@elponick That's great! yeah kiln is my solo side project just building it to learn stuff. Im trying to start something of my own too, still finding a problem statement good enough to work on.

[Rishi @RishivikramN]

@elponick Got it, are you the founder?

[Rishi @RishivikramN]

@elponick Will check for sure. Yeah likewise :)

[Rishi @RishivikramN]

@anuj_shashimal @X Im building a lightweight agent harness tool in public called kiln.

https://t.co/qmMBV5O31A

[Rishi @RishivikramN]

@beebomco Recently purchased Rs.80K variant, now it costs Rs.118900 :(

[Rishi @RishivikramN]

@elponick Sure, let stay connected, would love to share notes.

[Rishi @RishivikramN]

Architecture of the Kiln harness tool, will keep improvising it until Im confident to use it in one of my projects.

[Rishi @RishivikramN]

@paraschopra Currently, building a light-weight agent harness tool which can help me replace claude code in my workflow.

https://t.co/qmMBV5O31A

[Rishi @RishivikramN]

@elponick Yeah thats true. Yeah Im still working on it. Im currently using compact method for context management in multiple sessions. Will try to improve this forward. Where do you work btw?

[Rishi @RishivikramN]

Update: Added sessions to kiln. Right now, when I launch kiln in existing session it restores previous session's chat.

[Rishi @RishivikramN]

Im trying to implement kiln as a light weight tool which doesn’t have any bloat. Right now the binary is under 10Mb.

[Rishi @RishivikramN]

Built a simple proof of concept for an agent harness tool similar to claude code called kiln. currently, kiln perform basic tool calling such as list_files, read_files and write files. And I have added file diff render for writes.

[Rishi @RishivikramN]

I have also added permission management. Right now it is single agent loop with plain context feedback without any truncation.

[Rishi @RishivikramN]

@elponick Thats great, currently Im trying out with claude, gemini and qwen local models.

[Rishi @RishivikramN]

Very much in love with Claude code, although it does feels bloated. So Im writing my own agent harness which is lightweight and has only what is needed. will be posting more updates soon.

[Rishi @RishivikramN]

@elponick accuracy first,does the model still remember a fact from 10 turns ago or does it make something up? latency is next. 70% variance from chunk sizing is surprising though were you chunking tool outputs or the whole conversation history?