Waveflow: the harness for AI

Waveflow is the harness that makes AI effective for hardware design.

AI agents can generate HLS and drive design-space exploration — but they’re only as good as the substrate they work on, and raw HDL/HLS gives them none of what they need. Waveflow gives them four:

• Fast simulation — vectorized, orders of magnitude faster than RTL, so an agent can try many designs inside its loop instead of waiting on one toolchain run.
• Structured architecture — typed schemas and well-defined interfaces make code generation local: an agent fills in one component against an explicit interface contract, not a monolithic kernel it has to get entirely right at once. This is what lets AI scale past local fragments to whole systems.
• Deterministic, reproducible builds — the build graph runs the same way every time, so a generated design can be rebuilt, compared, and trusted.
• Built-in, bit-exact validation — every result is checkable against the real toolchain, so the output is verified, not just plausible.

Waveflow is that substrate.

AI is downstream, not the center

AI is a first-class downstream consumer of the representation — a real strength, and an active area of development — but it is grounded by the substrate, not the center of it. The codegen pipeline itself is deterministic (structured hwgen, not an LLM), and Waveflow is the substrate beneath an agent, not just an orchestration layer over one. That is the difference between AI output that is merely plausible and AI output you can trust.

AI in Waveflow (future work)

Waveflow is designed to integrate AI in three places — an active and largely forward-looking area of development:

• AI-assisted codegen. Generate a component’s HLS kernel from its Python model against the typed interface contract, so an agent fills in one well-scoped block at a time instead of a whole design. (Today the codegen pipeline is deterministic hwgen; AI assistance layers on top of it.)
• Agentic design-space exploration. An agent drives the fast inner loop, using tools that extract performance, timing, and resource estimates from the SimPy simulation to choose the next design point and iterate.
• MCP tooling. Model Context Protocol servers help users write Waveflow code and let AI agents query simulation and synthesis results directly — grounding suggestions in real artifacts rather than guesswork.

These map onto the Waveflow flow: AI assists at codegen and drives the agentic DSE loop, both grounded by the bit-exact substrate.