Every hot path, deliberately chosen

A drop-in replacement for LangGraph’s SQLite checkpointer. It moves the hot serialization path into Rust, running up to 737× faster than Python deepcopy on large state and 5–6× faster end-to-end.

Rust-backed state merging for high-frequency inner-loop updates — about 1.4 µs/call, and 45.9× faster than the Python path on tight 1,000-step workloads.

One import transparently patches LangGraph at load time. No graph code changes. Caches ThreadPoolExecutors across invocations, removing the largest single per-invocation cost.

Skip the monkey-patch and wire in individual Rust components exactly where you want them. Full control, same speedups, explicit call sites.

A one-line decorator around your LLM call site. Content-addressed by arguments, sub-microsecond lookup, configurable LRU eviction. 9.78× throughput at a 90% hit rate — and a direct cut to API spend.

The same cache works for expensive deterministic computations inside nodes — embeddings, parsing, scoring — for a ~1.6× speedup on repeated calls.

A low-overhead profiler that attributes wall-clock time across checkpointing, state updates, node execution, and LLM calls — so you accelerate the path that actually matters.

fast-langraph runs LangGraph’s own upstream test suite against its shimmed implementation — 85 pass, 0 fail, 3 skipped for environment reasons — against LangGraph 1.0.4.