Speeding up CrewAI: runtime vs. startup

Fast-CrewAI targets the runtime hot paths where Python overhead dominates: memory search is 11.2× faster via SQLite FTS5 + BM25 (vs. LIKE-query scans), JSON serialization is 34.5× faster via Rust serde, and repeated tool execution is 17.3× faster with result caching. These are per-subsystem microbenchmarks, not end-to-end numbers — real workflows typically see 1.3–5× overall speedups, with the biggest wins on memory- and database-heavy pipelines. Install it, add one import line, and your existing Agents, Tasks, and Crews use the accelerated paths automatically.

By default, CrewAI memory search relies on LIKE-query scans, which grow slower as stored memory grows because they scan rows rather than using a search index. Fast-CrewAI replaces this with a SQLite FTS5 full-text index and BM25 ranking, which is 11.2× faster than LIKE scans in our benchmarks. Memory- and database-heavy pipelines see the largest real-world benefit.

Yes. Fast-CrewAI is a drop-in Rust acceleration layer for CrewAI. It monkey-patches CrewAI components at import time and keeps 100% API compatibility, so you do not fork CrewAI or rewrite agent code — you add one import before your CrewAI imports and keep your existing Agents, Tasks, and Crews. It accelerates runtime execution (memory search, serialization, tool-result caching), not install or startup time.

CrewAI's own "100× Speed Boost" refers to startup: install, dependency resolution, and UV-based setup time — not the runtime cost of executing a crew. Fast-CrewAI addresses the other half: runtime execution. It accelerates memory search (11.2× via FTS5/BM25 vs. LIKE scans) and serialization (34.5× via Rust serde), plus tool-result caching (17.3×) and parallel scheduling with r2d2 connection pooling. These headline figures are per-subsystem microbenchmarks; end-to-end CrewAI workflows typically see 1.3–5× overall. So the two are complementary: CrewAI's 100× makes setup fast, Fast-CrewAI makes execution fast.