Research Summary
Prompt engineering has evolved from simple sentence inputs to a complex infrastructure mechanism directing agent identity, tool permissions, and system boundaries. As single-model techniques like zero-shot prompting and chain-of-thought are ported into multi-agent systems, new challenges arise.
This comprehensive survey analyzes over 35 works published between 2022 and 2026, revealing that prompt sensitivity propagates and amplifies in cascaded multi-agent architectures. The paper demonstrates that joint prompt-topology optimization is now a critical systems-level design problem, rather than an isolated tuning task.
Key Contributions:
- A unified taxonomy of prompt engineering techniques spanning foundational methods to multi-agent design patterns.
- Documentation of cascade amplification, where single-agent prompt sensitivity compounds across agent pipelines.
- Identification of three critical open challenges in production systems: constraint prompt ambiguity, cross-model portability, and the lack of multi-agent evaluation standards.
Research by eMitra Technologies
This research was co-authored by our Co-Founder Nishchay Gupta, alongside Prof. Shweta Jain and Prof. Anju Pawar. It represents our foundational approach to building scalable, robust, and secure multi-agent systems for our clients.
Read Full Paper