Scroll horizontally to explore the full diagram.
measure Collect
Supports the components above
Components of an agentic compiler harness
Select a component to inspect its role. The highlighted connections show how the agent uses it.
Agent workflow
The workflow coordinates optimization work. Agents retrieve knowledge, submit programs for analysis, run GPU experiments, and retain candidates and results to guide the next change.
Knowledge base
The agent retrieves complete kernels from the zoo and operation semantics from hardware specifications. Together they supply implementation ideas and the conditions needed to adapt them.
Domain-specific compiler analysis
The agent submits a kernel and receives diagnostics about its behavior. Reports identify the source operations and relationships involved so the agent can investigate and repair the implementation.
Minimal foundation
A small, stable program representation supports compilation and analysis. It exposes memory accesses, synchronization, and hardware operations, and can be extended as new primitives become available.
Benchmark server
The server runs the submitted correctness checks, benchmarks, and profiling captures on managed GPUs. It returns results and artifacts that the agent can use to compare candidates and investigate bottlenecks.
GPUs
The target devices execute the candidate kernels. The benchmark server coordinates their use so measurements from competing experiments do not interfere with one another.
Self-improvement across runs
Reviewing optimization results can expose gaps in tools and checks. Validated kernels add implementation knowledge to the zoo, improving the environment available to later agents.