A community effort to design computer architecture and computing systems with AI in the loop
Architecture 2.0 treats AI-assisted computer architecture as a community problem: shared design loops, tools, datasets, evidence, and review practices that make results easier to trust and easier to build on.
For fifty years, computer architecture advanced through design loops run by people: propose a change, model it, measure it, decide. Those loops now have to keep pace with trillion-transistor systems and workloads that shift faster than any team can hand-tune. The question Architecture 2.0 takes up is what happens when AI enters the loop - not to write a little code faster, but to explore, evaluate, and optimize designs directly.
Letting an agent inside the loop only works if the loop is trustworthy. Much of the effort here is about making design loops explicit and auditable: defining the state an agent reasons over, the actions it may take, the rejection criteria that disqualify a bad result, and the evidence that makes a claim believable. The synthesis lecture develops that vocabulary; the tool registry collects the simulators, models, and harnesses that make it runnable.
Architecture 2.0 is organized around the shared infrastructure a credible design loop needs.
No single lab can build this. Turning machine learning loose on architecture needs shared datasets, comparable benchmarks, open tools, and reproducible evidence - and it needs people trained to work across both fields. Architecture 2.0 grew out of our research at Harvard, but it is meant to be a commons that many groups build together, in the open. The work organizes into a few strands:
Shared, versioned data and comparable benchmarks so results mean the same thing across groups.
The learning, search, and agentic methods that generate, predict, and optimize designs inside a loop.
Open simulators, proxy models, and verification harnesses - the action space an agent operates in.
Practices that let anyone re-run a loop and check the evidence behind an architectural claim.
Courses and materials for a generation of architects fluent in both systems and machine learning.
The community behind this work has been meeting for years. Since 2020, the MLArchSys workshop at ISCA (the International Symposium on Computer Architecture) has brought the machine learning, systems, and architecture communities together around learning for hardware and hardware for learning - most recently in Tokyo (2025), and next in Raleigh (2026).
As foundation models and autonomous agents began to reshape how systems are designed, that community turned toward agentic approaches. MLArchSys 2026 adds a dedicated A³ (Agentic Approaches to Architecture) track, and the Architecture 2.0 workshop at ISCA 2026 focuses squarely on agentic AI for computing-systems design, anchored in computer architecture and hardware/software co-design. An early gathering, opened by a keynote from Partha Ranganathan (Google), helped map the datasets, tools, and training the field would need.
Architecture 2.0 itself grew out of Harvard's CS249r graduate seminar (Fall 2025) and the synthesis lecture developed alongside it. It is part of the mlsysbook.ai family of open, community-built learning resources.
Use the hub as the front door for reading, contributing, and teaching.
Built an open simulator, surrogate model, or agentic loop? Propose it for the registry.
Join the conversation in Discussions, and contribute on GitHub and the 🤗 Hugging Face artifact hub.
Adapt the book and course materials for your own students and reading groups.