TLDR: SLEAN (Simple Lightweight Ensemble Analysis Network) is a new framework designed to coordinate multiple Large Language Models (LLMs) for automated software bug investigation. It uses a three-phase protocol—independent analysis, cross-critique, and arbitration—to filter out harmful AI-generated code suggestions, ensuring reliable and auditable fixes without the complex infrastructure typically required by multi-agent systems. The research found that simpler bug descriptions (Type 2) led to more efficient fixes and that the arbitration phase was essential for preventing problematic suggestions from reaching production code.
In the rapidly evolving landscape of artificial intelligence, Large Language Models (LLMs) are becoming indispensable tools for tasks ranging from data analysis to code generation and automated program repair. However, their practical deployment often faces a dilemma: single models can be inconsistent and biased, while sophisticated multi-agent systems, though powerful, demand complex infrastructure and deep technical expertise.
Introducing SLEAN: A Lightweight Solution for LLM Coordination
A new framework called SLEAN, which stands for Simple Lightweight Ensemble Analysis Network, offers a pragmatic middle ground. Developed by Matheus J. T. Vargas from CyAura AI, SLEAN is designed to coordinate multiple LLM providers through a straightforward, text-based prompt orchestration. Unlike its more complex counterparts, SLEAN operates as a simple prompt bridge, utilizing basic .txt templates, making it accessible for deployment without requiring specialized technical knowledge.
The core of SLEAN is a three-phase protocol: independent analysis, cross-critique, and arbitration. This structured approach is crucial for filtering out potentially harmful AI-generated code suggestions before they are deployed into production. This addresses a growing concern in AI-assisted debugging, where modifications can sometimes introduce unnecessary complexity, break existing functionality, or fail to solve the actual problem.
How SLEAN Works: The Three Phases
SLEAN’s coordination protocol is built on three distinct phases that maintain independence while enabling structured synthesis:
- Phase 1: Independent Analysis: Each LLM provider receives the same task description and code corpus. They perform an independent audit without seeing the outputs of their peers. This isolation prevents early convergence and ensures diverse initial perspectives. The outputs are saved as artifacts, providing a baseline for review and supporting reproducible analysis.
- Phase 2: Cross-Critique and Consolidation: In this phase, each provider reads the original inputs, its own Phase 1 audit, and the peer’s Phase 1 audit. It then issues a consolidated assessment, highlighting points of agreement, disagreement, and proposed changes. This step facilitates peer review, exposing conflicts and potential errors while preserving each provider’s unique analysis.
- Phase 3: Synthesis and Arbitration: Finally, an arbitrating provider consumes all original inputs, both independent audits, and both consolidated assessments. This arbitrator then produces the definitive synthesis, listing accepted and rejected fixes with justifications and attributions. This sequential structure ensures determinism, minimizes uncontrolled interactions between models, and creates an auditable trail for continuous delivery pipelines.
Key Advantages Over Existing Systems
SLEAN distinguishes itself from other multi-agent frameworks like CrewAI, AutoGen, and LangGraph in several key areas:
- Deterministic Execution: Its fixed three-phase workflow ensures repeatable runs, which is vital for regression analysis in continuous integration/continuous delivery (CI/CD) environments.
- Provider Isolation: By mediating all interactions through an orchestrator and enforcing independence in Phase 1, SLEAN reduces correlated errors, a principle derived from ensemble learning.
- Controlled Aggregation: A single arbitrating provider makes definitive accept/reject decisions, providing a clear audit trail and lowering variance by combining diverse inputs.
- Provider Agnostic: SLEAN treats different LLM providers (e.g., OpenAI GPT, Anthropic Claude) as interchangeable analytical resources, leveraging their diversity without complex role negotiation.
- Full Auditability: It preserves complete artifacts at each phase, ensuring transparency and accountability, which is critical for production machine learning systems.
- Seamless CI/CD Integration: Its file-driven, configuration-based design eliminates infrastructure overhead, making it easy to integrate into existing development pipelines.
- Strict Scope Control: The arbitration process prioritizes minimal, causal edits, preventing over-engineering and scope creep often seen in conversational reasoning systems.
Case Study Findings: Efficiency and Risk Mitigation
The researchers evaluated SLEAN across 15 software bugs, analyzing 69 AI-generated fix propositions. The framework accepted 22 fixes (31.9% acceptance rate) and rejected 47 that would have been harmful if applied directly. The arbitration process significantly reduced the code change surface by 83-90% compared to raw AI outputs, emphasizing minimal, targeted interventions.
Interestingly, the study categorized user bug descriptions into two types: Type 1 (Technical/Diagnostic Issues with detailed information like stack traces) and Type 2 (Functional/Behavioral Issues with informal, colloquial descriptions). Type 2 inputs proved more efficient, requiring fewer propositions per accepted fix (2.85 vs. 3.56 for Type 1), representing a 20% efficiency gain. This suggests that sometimes, simpler problem statements can lead to more direct and efficient solutions from LLMs.
The analysis also revealed that one of the LLM providers (AIB, Anthropic Claude) consistently dominated in solution generation, outperforming the other (AIA, OpenAI GPT) in 14 out of 15 cases. While AI system agreement (convergence) showed a weak correlation with fix quality, the arbitration phase was crucial, contributing to 23% of the final solutions on average, and up to 50% in complex cases.
A critical aspect of SLEAN’s value is its ability to identify and reject risky AI-generated suggestions. These included fatal syntax errors, fixes for nonexistent problems, over-engineering, changes that would break existing functionality, misunderstandings of existing code, addressing symptoms instead of root causes, production code quality violations, and speculative enhancements beyond the scope of the bug.
Future Directions
While SLEAN offers significant advantages, the authors acknowledge limitations such as sequential processing (preventing parallel execution in some phases) and static provider assignment. Future work aims to explore parallel execution, adaptive provider selection based on performance, feedback integration, and hierarchical analysis for very large codebases. An important future enhancement is to use an independent third provider for arbitration to eliminate potential biases.
Also Read:
- Bridging the Communication Gap: Enhancing Robust Code Generation with AI Agents
- RA-Gen: A New Framework for Secure and Controllable Code Generation
Conclusion
SLEAN presents a robust and practical framework for coordinating multiple LLM providers in automated bug investigation. It achieves the analytical diversity of complex multi-agent systems without their inherent infrastructure complexity. By ensuring independence, supporting reproducible audits, and systematically filtering harmful proposals, SLEAN addresses critical safety concerns in AI-assisted development. Its provider-agnostic and file-driven design makes it broadly applicable to other domains requiring reliable, explainable, and auditable analytical consensus, such as security auditing, code review, and document verification. For more details, you can refer to the original research paper.
