AgentFactory: Towards Automated Agentic System Design and Optimization
A paper titled 'AgentFactory: Towards Automated Agentic System Design and Optimization' was published on arXiv (cs.AI) on 2026-09-01. The paper presents AgentFactory, a framework that jointly optimizes foundation models and workflow structures in agentic systems, considering multiple objectives including performance, cost, and efficiency. It uses LLMs as optimizers and a three-stage optimization pipeline.
The paper introduces AgentFactory, a framework for automated design and optimization of agentic systems. It addresses limitations of manual design and prior automated workflow optimization by jointly optimizing foundation models and workflow structures. The framework considers multiple objectives (performance, cost, efficiency) and uses LLMs as optimizers within a three-stage pipeline to explore configurations and discover effective combinations of fine-tuned models and optimized workflows.
AgentFactory employs a three-stage optimization pipeline where LLMs act as optimizers to navigate the search space of model and workflow configurations. The joint optimization of models and workflows, with multi-objective considerations, suggests a shift from single-metric workflow search to holistic system design. Observable next signals include empirical results on benchmark tasks, comparisons against manual and single-objective baselines, and details on the fine-tuning process.
The framework targets real-world deployment constraints by balancing performance, cost, and efficiency. This could reduce the manual engineering effort required to build agentic systems, potentially accelerating adoption in enterprise and developer contexts. The approach may influence tooling for automated agent design and optimization.
Automated optimization of agentic systems could lower development costs and time-to-market for AI agents. By jointly optimizing models and workflows, organizations may achieve better performance-cost trade-offs, making agentic solutions more viable for production use.
If validated, AgentFactory could lead to more adaptive and scalable agentic systems. Future work may extend the framework to additional objectives, dynamic environments, or broader model families. The paper's publication may prompt further research on automated co-optimization of models and workflows.