I previously shared this on Discord, but since it seems preferred to track feature requests here, I am reposting it on GitHub.
I first experienced parallel multi-agent functionality in Kimi Code, and it is also supported in frameworks like Antigravity. Once you try it, you realize how incredibly powerful it is. It entirely prevents passing a degraded context—where crucial information gets lost along a sequential pipeline—to subsequent modes.
You might ask, "Why not just use those other tools instead of bothering the Zoo Code maintainers?"
My answer is simple: Zoo Code is genuinely my favorite tool, and I have a lot of affection for this project. I want to see it become even better!
> Note: This feature request was structured and written with the assistance of AI to clearly and effectively articulate my thoughts.
Is your feature request related to a problem? Please describe.
Currently, Zoo Code’s orchestration functionality works exceptionally well. The Orchestrator flexibly invokes Ask, Debug, Code, and Architect modes based on the custom environment, reliably completing long working loops.
However, the current mode-switching mechanism fundamentally operates as a 'Serial Agent' system. This structure inherently limits the ability to gather and utilize diverse data independently before moving into the planning phase. The core of a robust plan relies heavily on how much high-quality data is gathered and what kind of resources are synthesized beforehand. In the current serial loop, each mode sequentially depends on the execution history and text context of the previous mode. This creates path dependency, accumulates context noise, and ultimately restricts the scope of exploration.
Describe the solution you'd like
I propose introducing 'Parallel Agent Orchestration' (a Fork-Join phase), allowing multiple modes to run simultaneously and independently to gather required data before transitioning to the planning and design phase (Architect Mode).
- Example Workflow:
- A complex user requirement is triggered.
- Before establishing a plan, the Orchestrator invokes [Research Mode], [Ask Mode], and [Debug Mode] simultaneously in parallel.
- These three modes independently gather and prepare three distinct, unbiased core contexts/datasets.
- [Architect Mode] receives these three independent data inputs all at once (Join/Reduce) to build and implement a highly sophisticated, comprehensive plan.
Implementing this architecture would deliver far more precise and well-rounded plans compared to a single linear loop.
Describe alternatives you've considered
- Sequential execution within the current structure: Attempting to gather data by running single modes multiple times sequentially. However, this linearly increases latency and causes the context window to accumulate previous loop histories, creating biases that hinder objective data collection.
- Reference Frameworks: Frameworks like Antigravity successfully employ a similar mechanism. They execute multiple agents in parallel to pre-collect and process data, then aggregate it at a higher phase to crystallize the plan. This approach has proven to deliver superior performance and architectural precision.
Additional context
Combining Zoo Code's highly stable orchestration loop with this parallel data aggregation capability would elevate the framework from a sequential code editor into a true next-generation multi-agent platform capable of handling complex software architecture design and multi-angle verification.
I previously shared this on Discord, but since it seems preferred to track feature requests here, I am reposting it on GitHub.
I first experienced parallel multi-agent functionality in Kimi Code, and it is also supported in frameworks like Antigravity. Once you try it, you realize how incredibly powerful it is. It entirely prevents passing a degraded context—where crucial information gets lost along a sequential pipeline—to subsequent modes.
You might ask, "Why not just use those other tools instead of bothering the Zoo Code maintainers?"
My answer is simple: Zoo Code is genuinely my favorite tool, and I have a lot of affection for this project. I want to see it become even better!
> Note: This feature request was structured and written with the assistance of AI to clearly and effectively articulate my thoughts.
Is your feature request related to a problem? Please describe.
Currently, Zoo Code’s orchestration functionality works exceptionally well. The Orchestrator flexibly invokes Ask, Debug, Code, and Architect modes based on the custom environment, reliably completing long working loops.
However, the current mode-switching mechanism fundamentally operates as a 'Serial Agent' system. This structure inherently limits the ability to gather and utilize diverse data independently before moving into the planning phase. The core of a robust plan relies heavily on how much high-quality data is gathered and what kind of resources are synthesized beforehand. In the current serial loop, each mode sequentially depends on the execution history and text context of the previous mode. This creates path dependency, accumulates context noise, and ultimately restricts the scope of exploration.
Describe the solution you'd like
I propose introducing 'Parallel Agent Orchestration' (a Fork-Join phase), allowing multiple modes to run simultaneously and independently to gather required data before transitioning to the planning and design phase (Architect Mode).
Implementing this architecture would deliver far more precise and well-rounded plans compared to a single linear loop.
Describe alternatives you've considered
Additional context
Combining Zoo Code's highly stable orchestration loop with this parallel data aggregation capability would elevate the framework from a sequential code editor into a true next-generation multi-agent platform capable of handling complex software architecture design and multi-angle verification.