Multi-agent mode

[maryasov]

Proposed Improvement: Implement a multi-agent development approach in Cline in which each agent is assigned a specific role (e.g., planning, searching for documentation on the libraries used, coding, code review, testing, and communicating with the user). This approach will allow for a more efficient distribution of tasks, allowing each agent to focus on its specialized role.
In addition, it should be possible to select the most appropriate model for a particular agent role. Models with reasoning, models with good coding skills, models with computer vision, etc.

Related discussions in the Cline repository:
cline#489
cline#1007
cline#1038
cline#387

[davidmoshal]

I'm wondering if this can be combined with the mcp-servers, each has a role, and I think each can use a different provider. Not sure if MCP servers can report back the token usage and cost, but presumably this is doable.

[mrubens]

I really like the idea of manually chosen "chat modes" as a first step. From there I think we can iterate on the right way to transition between the different modes, and eventually to the idea of the proxy that automatically decides the best mode. On that last point - I would love any thoughts people have on how that might work, or any good examples to look at. Thank you!

[dustmoo]

I love that, I'm using a pretty built out "Additional Prompt" but I don't want it to be the same personality all the time. Development, vs Test vs Documentation all have different fine-tuning and keeping that context optimal to manage expense is a big deal.

[DoiiarX]

I hope it too

[TomLucidor]

This might devolve into MetaGPT... or OpenHands when they have the new feature

[scottnzuk]

Agent to Agent Googles new A2A. be amazing if we can get it before anyone else lol.

[BradKML]

Why A2A over MCP tho? And why is everyone trying to make their own version of "MCP"?

[SmartManoj]

A2A != MCP

[BradKML]

Can't agents be treated as another "tool" and thus MCP is likely good enough for A2A usage? What are the key benefits?

[SmartManoj]

Agent-to-agent interactions are more dynamic, and frankly, a bit messy — just like human conversations.

[decentralize-all]

i need to build a bot which +1's @SmartManoj - dude, spot on :). Give that man a raise. (sorry for the notification to others though, could be considered noise lol).

[BradKML]

Issues with model remixing (and possibly model routing based on "regular" vs thinking/reasoning models) #3170 #503 #888

Multi-agent workflows either through parallel/collaborative agents, and async agents through "switching hats" #1999 #2985 #2007 #2978

@dosu please check these links out accordingly and aggregate their observations and derived suggestions

[dosubot[bot]]

Here’s an aggregated summary of the key observations and suggestions from the linked discussions on model remixing/routing and multi-agent workflows in Roo Code:

Model Remixing & Routing

• There is strong interest in enabling multiple models to collaborate on tasks, leveraging their unique strengths. For example, using OpenRouter to let models like o3 and Gemini 2.5 work together, iterating and providing feedback to each other, similar to a "Pilot and Copilot" setup. This could lead to more intelligent and robust outputs, but also raises concerns about increased token usage and the need for mechanisms (like cursor rules or MCPs) to manage these interactions efficiently 3170.
• The idea of an "LLM Router" is discussed, which would dynamically switch between models based on the nature of the task—using "regular" models for standard tasks and "reasoning" models for more complex, thought-intensive operations. This aligns with the broader goal of model remixing and routing to optimize both performance and cost 3170.
• A dual-LLM methodology is suggested, where agents use the same LLM with different prompts to create a dialectical process, allowing the model to challenge its own responses. This could be extended to use different LLMs for even richer interactions 503.
• Combining reasoning models with standard models is highlighted as a way to enhance multi-agent workflows. For example, using Deepseek R1 for reasoning and Claude 3.5 Sonnet for response generation, where the output of the reasoning model is used to improve the final response. This approach also opens up possibilities for dialectic argumentation between models 888.

Multi-Agent Workflows (Parallel/Collaborative & Async "Switching Hats")

• There is a clear push for integrating an "Agent Mode" in Roo Code, leveraging frameworks like Microsoft AutoGen. This would allow users to switch between "Classic Mode" and "Agent Mode," where multiple agents (planning, coding, testing, debugging) can work together autonomously. Key challenges include building a robust communication bridge between TypeScript and Python, and managing dependencies 1999.
• Asynchronous and parallel task execution is seen as essential, especially for optimizing time spent waiting for LLM server responses. One suggestion is to use "git worktree" to manage tasks in separate branches, enabling simultaneous work without threading or multiprocessing. This also hints at the need for collaborative agent functionalities similar to MetaGPT 2985.
• The potential for agent collaboration is recognized, but current implementations may not fully support it yet. There is ongoing interest in improving multi-agent functionalities, with users expressing a desire for agents to work together more seamlessly 2007.
• Defining distinct roles within the multi-agent workflow is emphasized. Suggested roles include Architects (system architecture), DevOps/Infrastructure (deployment docs), Developers (use cases for coding agents), and an Orchestrator (model selection and interaction management). There are concerns about avoiding conflicts, especially with API calls, and the need for structured workflows and clear role definitions 2978.

Common Themes & Suggestions

• Task specialization: Assigning specific roles to agents (planning, coding, testing, etc.) and allowing model selection per role to leverage different model strengths.
• Model orchestration: Implementing a proxy or router to automate model selection and task routing, reducing manual configuration.
• Collaboration and dialectics: Encouraging agents/models to challenge and refine each other's outputs for higher quality results.
• Workflow structure: Emphasizing the importance of clear role definitions, structured workflows, and mechanisms to prevent conflicts in multi-agent setups.
• Technical challenges: Addressing increased token usage, managing dependencies, and building robust communication bridges between components.

These discussions collectively point toward a future where Roo Code supports highly flexible, collaborative, and efficient multi-agent workflows, with intelligent model routing and remixing at their core.

To reply, just mention my name @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}