FHE-Agent: Automating CKKS Configuration for Practical Encrypted Inference via an LLM-Guided Agentic Framework

FHE-Agent addresses a critical bottleneck in privacy-preserving MLaaS: the 'Expertise Gap.' Configuring FHE parameters like the CKKS scheme's RNS levels and scale is notoriously difficult even for cryptographers. While the project has 0 stars and 6 forks, indicating it is currently a niche academic research artifact (linked to arXiv 2511.18653v1), its approach—using LLMs as heuristic engines for complex configuration spaces—is a growing trend. Its defensibility is currently low (3) because it lacks a community, production-grade API, or integration with mainstream FHE compilers like Zama's Concrete or Google's FHE-C++. The moat would depend on the quality of its internal evaluators and feedback loops, which are currently just a research prototype. Frontier labs like OpenAI or Anthropic are unlikely to build this directly as they focus on general intelligence, but infrastructure players like Microsoft (SEAL) or IBM would likely absorb such capabilities into their own toolchains if the technique proves superior to traditional optimization algorithms like Simulated Annealing or Bayesian Optimization. The project's value lies in its 'novel combination' of agentic workflows with cryptographic parameter search, but it risks displacement by dedicated FHE compilers that are increasingly using ML for auto-tuning.