Fast-CrewAI

Features

What Fast-CrewAI accelerates, and by how much

Fast-CrewAI targets the runtime hot paths where Python overhead dominates a CrewAI run. Every number below comes from the public benchmark suite against a pinned CrewAI version — per-subsystem microbenchmarks, not end-to-end claims.

Rust serde serialization

CrewAI passes JSON payloads between agents, memory, and tools constantly. Fast-CrewAI routes that hot path through serde via PyO3 instead of Python's json module.

34.5× faster · 58% less memory

FTS5 memory search

Default CrewAI memory retrieval uses LIKE scans that slow down as memory grows. Fast-CrewAI swaps in a SQLite FTS5 full-text index with BM25 ranking.

11.2× faster · 31% less memory

Tool result caching

Agents frequently re-invoke the same tool with the same arguments. Fast-CrewAI caches results and validates with serde, collapsing repeated calls.

17.3× faster · 99% less memory

Pooled DB connections

Concurrent agents contend for SQLite connections. Fast-CrewAI uses r2d2 connection pooling so parallel access no longer serializes on a single handle.

1.3× faster DB queries

100% API compatibility

No fork, no rewrite. Fast-CrewAI monkey-patches CrewAI at import time via dynamic inheritance. Your Agents, Tasks, and Crews keep their exact API.

101 compatibility tests passing

One-import activation

Add import fast_crewai.shim before your CrewAI imports and the accelerated paths switch on automatically. Nothing else in your codebase changes.

Zero code changes

Granular toggles

Disable acceleration globally with FAST_CREWAI_ACCELERATION=0 for debugging, or turn individual accelerated components on and off independently.

Env-var controlled

Safe Python fallback

Pre-built wheels ship for Linux, macOS, and Windows on x86_64 and ARM64. If a native wheel is unavailable, Fast-CrewAI falls back to pure Python.

Cross-platform wheels

Open source, MIT

The full acceleration layer is MIT-licensed and developed in the open, with a public benchmark suite gating every release against a pinned CrewAI version.

Reproducible benchmarks

Feature deep dives

Serialization

Rust serde in place of Python json

34.5× · 80,525 vs 2,333 ops/s · 58% less memory

CrewAI serializes agent messages, memory records, and tool results as JSON on nearly every internal hop. Fast-CrewAI patches that path to encode and decode with serde through PyO3. On a representative agent payload the benchmark suite measures 80,525 ops/s versus 2,333 ops/s for the Python stdlib — a 34.5× improvement with 58% lower peak memory.

Memory search

SQLite FTS5 + BM25 in place of LIKE scans

11.2× · 10,206 vs 913 ops/s · 31% less memory

Default CrewAI memory retrieval uses LIKE-query scans that walk stored rows, so latency rises as memory grows. Fast-CrewAI builds a SQLite FTS5 full-text index and ranks with BM25. The result is 11.2× faster search that stays fast as the memory store expands — the single biggest win for long-running, memory-heavy crews.

Tool execution

Result caching with serde validation

17.3× · 11,616 vs 670 ops/s · 99% less memory

Agents routinely re-invoke the same tool with identical arguments. Fast-CrewAI caches tool results and validates payloads with serde, so repeated calls short-circuit instead of paying full execution cost. The benchmark records 17.3× higher throughput with 99% less memory on this hot path.

Database

r2d2 connection pooling

1.3× on concurrent DB queries

Concurrent agents contend for a single SQLite connection, serializing work that should overlap. Fast-CrewAI introduces r2d2 connection pooling so each worker gets its own handle, lifting concurrent database throughput by roughly 1.3× in the benchmark suite.

Compatibility

100% CrewAI API, verified by 101 tests

101 compatibility tests passing against crewai==1.7.2

Fast-CrewAI monkey-patches CrewAI components at import time using dynamic inheritance, so the public API is byte-for-byte the CrewAI you already use. A suite of 101 compatibility tests runs against a pinned crewai==1.7.2 on every release to guarantee behaviour matches upstream.

Activation

One import, zero refactor

import fast_crewai.shim

Add import fast_crewai.shim before your CrewAI imports and every accelerated path switches on. There is no configuration file to author and no code to migrate — your existing Agents, Tasks, and Crews immediately run through the faster implementations.

Controls

Granular acceleration toggles

FAST_CREWAI_ACCELERATION=0

For debugging or A/B measurement you can disable acceleration globally with FAST_CREWAI_ACCELERATION=0, or toggle individual accelerated components independently. That makes it trivial to isolate a regression to the framework versus the acceleration layer.

Portability

Cross-platform wheels with Python fallback

Linux · macOS · Windows · x86_64 & ARM64

Pre-built wheels ship for Linux, macOS, and Windows on both x86_64 and ARM64. When a native wheel is not available for a platform, Fast-CrewAI gracefully falls back to pure Python so installs never break.

Open source

MIT-licensed, reproducible benchmarks

Public benchmark suite gates every release

The whole acceleration layer is MIT-licensed and developed in the open. Every published number comes from the public benchmark_test/ suite, run against a pinned CrewAI version — if you cannot reproduce it, it does not ship.

Want the full methodology? Read the benchmark deep dive, see the architecture, or compare against stock CrewAI.

Ready to make CrewAI faster?

Talk to the team that wrote the acceleration layer. We take on performance audits, full system builds, and retained engineering.