Knowledge Base: First Community Contribution

New entry

• 034 — Optimizing Label Propagation in Graph Clustering: Full AAE case study on KaHIP/ScalableCorrelationClustering. 30 experiments achieving 1.23x speedup (18.7%) via dense vectors replacing hash maps, counting-sort contraction, LP sweep specialization, persistent buffer reuse, and tcmalloc. Includes detailed experiment log (18 kept, 12 discarded) and what-didn't-work analysis.

Highlights

• First knowledge base entry contributed from a real AAE session
• Covers multilevel graph framework optimization end-to-end
• Documents both successful and failed hypotheses with root-cause explanations

Knowledge Base: 30 Performance Optimization Entries

Seed the knowledge base with 30 expert entries covering C++ and Python performance optimization techniques.

C++ (entries 004–018)

• Cache efficiency: SoA vs AoS, hot-cold splitting, loop tiling
• Branch & instruction optimization: branchless programming, SIMD vectorization, compiler intrinsics
• Memory management: small buffer optimization, move semantics, reserve/preallocate, mmap, arena-style hash maps
• Compiler techniques: constexpr, PGO + LTO
• Concurrency: false sharing avoidance
• API design: std::string_view

Python (entries 019–033)

• Vectorization & JIT: NumPy vectorization, Numba JIT, Cython
• Memory reduction: slots, generators, array module, mmap
• Data structures: dict/set O(1) lookup, deque for queues
• Patterns: preallocation, string join, local variable caching, itertools pipelines, struct packing
• Parallelism: multiprocessing to bypass GIL

What's New

Knowledge Base

• Community-contributed knowledge base for sharing optimization techniques, experiment results, and lessons learned from AAE sessions
• Agents can fetch and search the knowledge base from GitHub to inform hypothesis generation
• Contributors submit entries via PR (one entry per PR, with INDEX.md update)
• Three example entries included (blocked sorting, gradient accumulation, arena allocation)

AAE Program Enhancements

• User assertions: The agent now formulates correctness invariants from the code and presents them to the user for confirmation during setup. Assertions are checked after every change, before benchmarking.
• Subcomponent replacement: Explicitly allows replacing entire algorithms, data structures, or modules (not just incremental tuning) when the analysis supports it.

Housekeeping

• Moved image files to img/ folder
• Added knowledge base banner image

• Add Agent Teams section for keeping the AAE loop alive
• Encourage GNU parallel for multi-run experiments
• Use geometric mean as default aggregation statistic

Full Changelog: https://github.com/CHSZLab/AgenticAlgorithmEngineering/commits/v1.0