jwbron

I'm using Claude code + (a patched) litellm proxy + openrouter + Qwen 3.7 max/kimi k2.6/deepseek v4 pro. The only feature that doesn't work is webfetch and web search, which I've replaced with the ddg MCP and a web fetch/search pre hook to redirect the agent. Memory, caching, and everything else works fine.

Qwen comes close to opus for planning but fable is clearly superior. Results for kimi and deepseek are pretty much indistinguishable from opus for coding if opus writes the plan. The biggest difference is output cadence. Kimi for example thinks for a long time then quickly outputs a lot of text.

I'm now testing out fable for research and planning and deepseek v4 flash for coding. I'm guessing results will be pretty similar to opus + deepseek v4 pro and costs should be lower overall.

I'm using Claude code + (a patched) litellm proxy + openrouter + Qwen 3.7 max/kimi k2.6/deepseek v4 pro. The only feature that doesn't work is webfetch and web search, which I've replaced with the ddg MCP. Memory, caching, and everything else works fine.

Qwen comes close to opus for planning but fable is clearly superior. Kimi and deepseek are pretty much indistinguishable from opus for coding if opus writes the plan.

I'm now testing out fable for research and planning and deepseek v4 flash for coding. I'm guessing results will be pretty similar to opus + deepseek v4 pro and costs should be lower overall.

The former GitHub CEO has a startup and this was their first release. They call it checkpoints: https://entire.io/

I copied it for my own tooling to make it work a bit better for my workflows.

Important note: this thing eats tokens. I'm currently cycling through four different claude max 20x accounts, hitting weekly limits every couple of days.

Egg is a side project of mine. It's a structurally enforced SDLC pipeline for autonomous LLM agents. You provide it with a task, it:refines the task definition and sends it back to you, you review the refinement, it generates a plan based on the refined task, you review the plan, it implements the plan, validates it using a contract validation mechanism, and opens a PR in github

It's comprised of what I'm calling "agentic feedback loops" https://github.com/jwbron/egg/blob/main/docs/agentic-feedbac... and "human feedback loops". There are three phases of "agentic feedback loops", which do work, review the work, and repeat until consensus is gained: refine, plan, implement (this also has an automatic merge conflict fixer, can run checks like test and lint, and has an autofixer. The contract is enforced in this stage)

Each phase has a human checkpoint in between. The human reviews the work, makes manual changes or iterates with claude, then approves. The final check is a human PR review. I'm finding that if I focus energy on refining the initial task, I always approve the review on the first pass. If the requirements are properly defined, they are enforced in the contract. This is managed by phase-based access controls. Coders can't alter the contract, contract validators can't modify the code. Agents must gain consensus. I call this the "SDLC pipeline" https://github.com/jwbron/egg/blob/main/docs/guides/sdlc-pip....

There's a lot going on here around sandboxing and enforcement mechanisms. This is covered more here: https://github.com/jwbron/egg/blob/main/gateway/README.md, but the documentation is slightly out of date. I'm actively working on an automated documentation intelligence system that ensures everything is up to date and should have that working in an hour or to.

I'd love to talk about this with folks. A lot of startups are announcing millions of dollars in funding to build out this kind of tool, but the reality is anyone can do it right now. I want to keep this open source and provide it to everyone for free, so we're all enabled to move as fast as we want with AI tooling. Hope I can build this out in a way that a lot of people find useful. This is early stages now, but I'm stoked to see how it continues to come together.

Right, but paying for food at a restaurant doesn't get me any closer to owning a restaurant. If promises are delivered on for these agentic coding platforms (which I do believe in), it seems the most reasonable path forward is to use those platforms to build your own platform.

The software collects the network traffic of distributed Claude code instances.

I built a basic copy with about an hour with my own "platform for ai agents" that I built out over the last week: https://github.com/jwbron/egg/pull/504, and refined it here: https://github.com/jwbron/egg/pull/517 (though right after I merged this I blew through my weekly token quota for my second claude max 20x account so I haven't been able to test it out yet).

I think your point is valid and I've been having the same thoughts. My tooling is still in the experimental phase, but I can move so quickly that I'm having trouble grasping how products like this will survive. If I can build this out in a week and copy an idea like this one (which is a great one, mind you) in an hour, what's the value of paying someone for a product like this vs just building it myself?