Every agent can complete any microtask successfully. At the same time, every agent will fail once complexity of scope and context gets too high, and it has to navigate planning, implementation and review on different abstraction levels. We realized this and built zeroshot, where planning, implementation and feedback is delegated to isolated agents. The feedback agents have non-negotiable veto powers over the worker agent, meaning that the usual tendency of optimizing for completeness over correctness gets avoided. We are using zeroshot to build a 1 million LOC hosted platform, and due to our trust in the tool to produce senior-level, scalable code, we have completely stopped looking at the codebase.
Author here. Zeroshot spawns isolated AI agents that work in parallel and with non-negotiable feedback loops. It ensures robust vibecoding. We're now adding support for codex and gemini.
The problem with current approaches is the lack of feedback loops with independent validators that never lose track of the acceptance criteria. That's the next level that will truly allow no-babysitting implementatons that are feature complete and production grade. Check out this repo that offers that: https://github.com/covibes/zeroshot/
The problem with current approaches is the lack of feedback loops with independent validators that never lose track of the acceptance criteria. That's the next level that will truly allow no-babysitting implementatons that are feature complete and production grade. Check out this repo that offers that: https://github.com/covibes/zeroshot/
The problem with current approaches is the lack of feedback loops with independent validators that never lose track of the acceptance criteria. That's the next level that will truly allow no-babysitting implementatons that are feature complete and production grade. Check out this repo that offers that: https://github.com/covibes/zeroshot/
The problem with current approaches is the lack of feedback loops with independent validators that never lose track of the acceptance criteria. That's the next level that will truly allow no-babysitting implementatons that are feature complete and production grade. Check out this repo that offers that: https://github.com/covibes/zeroshot/
zeroshot is a CLI that orchestrates autonomous agent clusters with feedback loops to ensure that the outputs are always feature complete and production-grade. Basically robust vibecoding with no human expert in the loop. Since launch we've had more than 1000 npm installs and 78 GitHub stars.
We’ve long been frustrated that, despite being insanely powerful, AI agents need a lot of handholding to arrive at a robust solution to the task you give it. Ralph Wiggum agents is the naive solution to this, but we’ve found that the need for this handholding completely disappears when independent review agents are tasked with validating the agents’ work.
So we build this open source tool on top of Claude code (will extend to other AIs soon!) that spawns a cluster of agents that operate together. They all have different roles, and can be triggered by different events. The framework is completely flexible and you can define any cluster configuration and collaboration structure among the agents that you want. However, the defaults should be quite powerful for most software dev.
By default, There’s also a routing agent (the conductor) tasked with classifying the task and deciding the optimal cluster for solving it. The clusters then usually accomplish the task to completion without any shortcuts, and key to this are the dedicated planning agents and independent review agents with separate mendates and veto power.
It’s easy to install and works out of the box with Claude code! Feel free to make issues if there you have improvement ideas or feature requests - we iterate very fast!