LLM Agents Factory: 20K prebuilt agent profiles slash inference costs

Deploying LLM agents in production usually means generating custom agent designs for each request—a process that's computationally expensive and unstable. The new paper 'LLM Agents Factory: Retrieval of Domain-Specific LLM Agents' from Vitalii Belov and colleagues at AIRI (Artificial Intelligence Research Institute) flips this paradigm: instead of generating agents from scratch, they retrieve them from a structured repository of over 20,000 pre-built agent profiles. The framework supports two modes: semantic search over these profiles to find the best-fitting agent, or distillation into a compact fine-tuned model that can generate the agent directly. This 'retrieval instead of construction' approach makes agent design fast, controllable, and reproducible, without sacrificing task quality.

In experiments across MMLU, BIG-bench, and BIG-bench Hard in single-agent scenarios, LLM Agents Factory outperformed standard non-agent baselines in accuracy and matched the generation quality of AutoGen—a well-known dynamic agent framework—when AutoGen used a 120B parameter backbone. The key advantage: LLM Agents Factory achieves this parity with dramatically lower inference cost, since it sidesteps expensive on-the-fly generation. The paper, accepted to SIGIR 2026, includes open-source code and the full agent profile base. For engineering teams, this means domain-specific agents can now be deployed at scale with predictable performance and budget, answering the strict demands of real-world industrial applications.