Agentic Architect: An Agentic AI Framework for Architecture Design Exploration and Optimization

2026-04-29 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Computer Architecture · Depth: Expert, extended

Summary

Agentic Architect is an open-source agentic AI framework developed by researchers at Carnegie Mellon University in 2026 for computer architecture design exploration and optimization. It integrates Large Language Model (LLM)-driven code evolution with cycle-accurate simulation to automatically generate and refine microarchitectural designs. The framework was evaluated across three critical domains: cache replacement, data prefetching, and branch prediction. Agentic Architect achieved significant performance improvements, including a 1.062x geomean IPC speedup over LRU for cache replacement (0.6% over Mockingjay), a 1.76x geomean IPC speedup over no prefetching (17% over VA/AMPM Lite and 21% over SMS) for prefetching, and a 1.100x geomean IPC speedup over Bimodal for branch prediction (1.5% over Hashed Perceptron). The framework's modularity allows interchangeability of LLMs, evolutionary frameworks, and simulators, and its analysis highlights the human architect's crucial role in defining search parameters like seed design, scoring functions, and prompt strategies.

Key takeaway

For AI Engineers and Research Scientists focused on microarchitecture design, Agentic Architect offers a powerful co-design paradigm. You should focus on defining robust search parameters, including strong seed policies, precise scoring functions, and diverse training workloads, as these critically bound the LLM's ability to discover high-performing, generalizable designs. Your expertise in structuring the search space will yield more effective and reliable architectural innovations.

Key insights

LLM-driven code evolution combined with cycle-accurate simulation can discover novel, high-performing microarchitectural designs.

Principles

• Evolution refines strong seeds, but cannot compensate for weak foundations.
• Minimal prompts outperform prescriptive prompts in LLM-driven evolution.
• Diversity in training traces is crucial for generalization.

Method

Agentic Architect uses an evolutionary loop where an LLM mutates code, a simulator evaluates performance, and a scoring function guides selection, with human input defining the search space.

In practice

• Use cost-effective LLMs for simpler evolution tasks.
• Prioritize diverse access patterns in training trace selection.
• Incorporate storage as an explicit optimization objective.

Topics

• Agentic Architect
• LLM-driven Code Evolution
• Microarchitectural Optimization
• Cache Replacement Policies
• Data Prefetching

Code references

• algorithmicsuperintelligence/openevolve
• MoonshotAI/Kimi-K2.5

Best for: AI Scientist, Research Scientist, AI Engineer

Related on AIssential

• ChatNeuroSim: An LLM Agent Framework for Automated Compute-in-Memory Accelerator Deployment and Optimization — ChatNeuroSim automates CIM accelerator design and optimization using LLM agents and design space pruning for faster, more efficient exploration.
• Agentic Discovery of Neural Architectures: AIRA-Compose and AIRA-Design — LLM agents can autonomously discover and optimize neural architectures and training methods, surpassing human-designed baselines.
• Agentic System Design Interview Series #1: Design a Coding Agent — Agentic system design requires building deterministic frameworks around non-deterministic LLM cores, emphasizing feedback loops and robust verification.
• From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms — LLM agent memory evolves through stages of storage, reflection, and experience, driven by consistency, dynamic environments, and continual learning.
• HyEvo for Topological Reasoning Graph Optimization — A self-evolving system autonomously designs optimal hybrid LLM-code agent workflows, eliminating manual architecture design.
• Beyond Correctness: Enhancing Architectural Reasoning in Code LLMs via Scalable Labeling with Agentic Judgment — Agentic LLM judges offer a scalable solution for evaluating and enhancing architectural reasoning in code generation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.