A3: Agentic Approaches to Architecture

First International Workshop on Agentic Approaches to Computer Architecture

Co-located with MICRO 2026, on November 1. A forum for exploring how reasoning-capable AI agents redefine the design of computer architectures and systems.

A³ workshop logo

A paradigm shift for computer architecture

The rise of reasoning-capable AI marks a paradigm shift for many design tasks, including computer architecture and systems design. Autonomous AI agents are now capable of completing long-horizon tasks, such as multi-step planning, algorithmic design, and tool manipulation, that were once carried out by expert human architects. The A3 (Agentic Approaches to Architecture) workshop explores how this evolution redefines the design of computer systems, exploring new AI-driven methodologies.

The objective of A3 is to catalyze a community around agentic methodologies for computer architecture. The workshop will strive to become a forum for presenting early-stage work in this nascent area, and synthesizing best practices through panel discussions and expert talks.

What the workshop explores

As a focused subset of the broader ML-for-Systems space, A3 centers on agentic methodologies for computer architecture. These workflows incorporate the following aspects.

01

Perception (Environment Sensing)

Real-time monitoring of architectural and system telemetry, from cache-miss distributions and OS resources to thermal and power envelopes.

02

Reasoning & Planning

Moving beyond "if-then" logic to dynamic planning. Agents use logic engines (LLMs or RL) to decompose goals (e.g., "minimize latency under a 10W constraint") into actionable sequences.

03

Tool Use (Action)

Agents exerting direct control, such as calling EDA APIs, triggering RTL simulations, managing architectural simulations, operating-system support, or reconfiguring hardware registers in situ.

04

Memory & Iterative Refinement

Systems maintain a "log" of outcomes to avoid past failures. If an action yields sub-optimal results, the agent re-reasons and pivots its strategy.

05

Architecture & Systems Agentic Co-design

Architectures and systems that are agent-facing, including abstractions, interfaces, and programmability requirements that cut across microarchitecture, computer architecture, operating systems, and VLSI.

Two ways to get involved!

Tentative workshop schedule

This will be a half-day workshop featuring the following sessions:

Organizers

A3 is jointly organized by Google/Google DeepMind, UC Berkeley, Carnegie Mellon University, and NVIDIA. The organizers can be reached at: a3-workshop@eecs.berkeley.edu

Google UC Berkeley UC Berkeley Carnegie Mellon University NVIDIA NVIDIA
Akanksha Jain
Google
Sagar Karandikar
UC Berkeley
Dimitrios Skarlatos
Carnegie Mellon University
Amir Yazdanbakhsh
Google DeepMind

Sponsors

Thanks to Google for sponsoring A3 and the A3 CHIA hackathon!

Google