New Framework Enhances Prompt Adaptation in Multi-LLM Agent Systems

Context adaptation is a technique that automates prompt engineering in Large Language Model systems by iteratively refining prompts based on task feedback, without altering the model's core weights. Extending this capability to multi-LLM agentic systems, where multiple LLMs collaborate on a task, presents challenges, particularly in accurately attributing performance to individual agents and guaranteeing the stability of prompt updates. This research proposes a new framework called Graph-based Target Back-Propagation (GTBP). GTBP models agentic workflows as directed acyclic graphs, allowing it to propagate local target outputs backward through the graph. This mechanism uses discrepancies between target and actual outputs to guide a stage-wise prompt update process. The theoretical analysis demonstrates that GTBP's prompt updates stabilize over iterations and that a capable LLM optimizer can reduce the overall objective. Empirically, GTBP consistently surpasses other strong baseline methods across various benchmarks, all while maintaining comparable computational efficiency, making it a robust solution for complex multi-agent AI systems.