I'm building an IDE (www.kaiso.ai)
AI is powerful, but currently does not meet the engineering bar for quality and thoroughness. We need new paradigms and tools to support a new relationship with the codebase as an artifact.
The premise is that we can use these LLMs to get real engineering work done if we make tools to support a higher-level human understanding of the codebase, and the ability to spot the gaps in the LLM's plans. With these we can surgically ensure all the critical considerations are covered, spec the work at an incredibly granular level, and implement our plans as a collection of ultra-tiny tasks each given to isolated agents, this specifically ensures the agent's attentional mechanism aren't overwhelmed/polluted.
The project is very early still, so if you're interested, please reach out or signup for the email-list and i'll contact you. Pricing page is highly aspirational at the moment, money is not the focus at this phase.
Thanks.
> The project is very early still, so if you're interested, please reach out or signup for the email-list and i'll contact you. Pricing page is highly aspirational at the moment, money is not the focus at this phase.
Why do you think an IDE is the right tool?
I'm working in a similar space, and it's not clear why an IDE would benefit.
Specifically to you - if you're hoping to make this a business - please know if you do make a killer IDE feature - Cursor et al will immediately copy it...
I'll give your tool a try if it's not too much effort to try it and you want some feedback. Let me know.
> Why do you think an IDE is the right tool?
I didn't start with an IDE but ended up there after some time. The core of my approach is an entirely new workflow. Underlying all of it is a "planning canvas" which is a network graph visualization of the codebase symbols, structures, and relations, where each node of the graph is a custom data-structure that captures a set of considerations. The workflow is generally as follows: Talk to the agent -> Agent responds with a plan(s) -> Plan is visualized on the planning canvas. At this point we can see visually which parts of the codebase the agents plan touches and via the fields of the custom data-structure, also see which considerations the agent failed to specify. Its here where we as humans can catch "this thing isnt connected, or is missing a trigger, or has a concurrency story, etc.", and either specify ourself, or force the agent to improve their plan in this specific manner. Once satisfied, we can formalize the impoved plan into a spec-of-specs, where each isolated sub-spec is farmed to an agent for implementation, which undo/redo being handled at the plan-level just in case we change our minds.
> Cursor et al will immediately copy it...
This is always possible, with anything and everything, but thus far they havent done it and i want this to exist, so i persist.
> I'll give your tool a try if it's not too much effort to try it
If you're open to it, signup (so i have your email) and ill reach out to get us going.
> Underlying all of it is a "planning canvas" which is a network graph visualization of the codebase symbols, structures, and relations, where each node of the graph is a custom data-structure that captures a set of considerations.
Cool, I'm thinking along the same lines.
> but thus far they havent done it and i want this to exist, so i persist.
Cool, we are in the same boat [=
> If you're open to it, signup
I'll check it out.