Argue: A Framework-Agnostic Multi-Agent Consensus Workflow Orchestration Solution

Argue is a multi-agent consensus workflow orchestration package decoupled from specific AI frameworks. It supports different agents to reach consensus through debate to enhance the reliability of AI decisions. This solution is framework-agnostic, compatible with various AI models and tools, has wide application scenarios, and focuses on modularity and composability in design, providing additional security guarantees for AI systems.

Single large language models tend to produce inconsistent outputs in decision-making scenarios due to training data biases, prompt changes, or randomness, which is unacceptable in critical scenarios. The way humans reduce individual biases by reaching consensus through discussion, debate, and voting has inspired AI systems. The Argue project introduces this idea into AI, allowing multi-agents to enhance reliability through 'debate'.

Argue is a harness-agnostic orchestration package that does not bind to specific AI frameworks (such as OpenAI GPT, Anthropic Claude, open-source Llama) or tools (LangChain, LlamaIndex, etc.). This design reflects the philosophy of the author onevcat (Wang Wei): to provide focused, composable tools rather than large, all-encompassing closed solutions.

1. Multi-Perspective Generation

When facing a decision problem, multiple agents independently output opinions. Different system prompts can be configured to simulate expert roles (such as auditor, entrepreneur, technical expert).

2. Structured Debate

After initial opinions are generated, agents review each other's views and put forward supporting or opposing arguments, simulating human expert group discussions to expose problems and blind spots.

3. Consensus Reaching

After multiple rounds of debate, extract the intersection of views to form the final consensus; if full agreement cannot be reached, output the majority opinion and mark the differences for human reference.

Code Review and Bug Detection

Multi-agents analyze code from the perspectives of security, performance, and maintainability, discovering issues missed by a single reviewer.

Content Moderation and Fact-Checking

Reduce misjudgments and balance strict and lenient judgment standards.

Complex Decision Support

In high-risk scenarios such as business decisions and medical diagnosis, reduce the impact of individual agents' 'hallucinations'.

Creative Generation and Evaluation

Inspire creativity and screen the most valuable ideas.

• Modular Architecture: Each stage of the consensus algorithm can be independently configured and replaced
• Asynchronous Execution Support: Parallel execution of agent debates improves efficiency
• Observability: Detailed intermediate state output facilitates debugging and auditing
• Cost Control: Dynamically adjust the number of debate rounds based on confidence to reduce unnecessary API calls

Compared to single-agent solutions, Argue provides higher reliability; compared to simple voting mechanisms, the debate process generates more intermediate insights; compared to other multi-agent frameworks, its framework-agnostic nature makes it easier to integrate into existing systems.

Argue represents a design paradigm of multi-agent collaboration + explicit consensus mechanism to enhance output reliability, drawing on human collective wisdom to provide a security layer for AI applications. As AI applications expand in high-risk domains, such consensus mechanisms will become standard practice.