Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-05 · high
Model checking spanned 7.4 million states before crash tests caught concurrent resumes firing one effect twice or more.
Why now: The paper carries an arXiv identifier placing it in August 2026, testing pinned releases of all three: LangGraph 1.2.9, CrewAI 1.15.2 and pydantic-graph 1.x.
LangGraph, CrewAI and pydantic-graph each violate a six-property resume contract that researchers model-checked across 7.4 million states, per a paper posted to arXiv in August 2026.
That gap isn't abstract for anyone running these frameworks with money or customer emails behind a paused step.
The paper, "Resume Means Resume" (arXiv 2608.03836), builds an LLM-free harness that fault-injects pinned releases and checks what each framework does against its resume contract.
Of the three, LangGraph 1.2.9 racks up the most failure modes. It durably records a second resume value and never reads it back. It writes schema-invalid state to disk without raising an error. After a real SIGKILL, it re-executes work it already durably recorded. The same API skips re-execution cleanly after a plain interrupt, so it's exactly-once for interrupts and at-least-once for crashes.
CrewAI 1.15.2 re-executes completed, effect-bearing methods, directly contradicting its documented resume behavior. pydantic-graph 1.x is more honest about its limits: it simply can't resume after a crash mid-node.
None of the three frameworks share a conformance profile with each other. The failure compounds under concurrency. When multiple processes resume the same parked interrupt at once, the paper found the gated effect fires once per process instead of once total. That saturated 36 of the 40 fault-injection cells tested.
Each link below shares sources, entities, or timing with this story.
LangGraph benchmarked against CrewAI / Shared entities / Earlier coverage
Linked by a graph relationship (LangGraph benchmarked against CrewAI); both cover CrewAI, LangGraph, LLM; earlier CrewAI coverage from 2026-06-29.
CrewAI supports Model Context Protocol / Shared entities / Earlier coverage
Linked by a graph relationship (CrewAI supports Model Context Protocol); both cover CrewAI, LLM; earlier CrewAI coverage from 2026-06-17.
LangGraph benchmarked against CrewAI / Shared entities / Earlier coverage
Linked by a graph relationship (LangGraph benchmarked against CrewAI); both cover CrewAI, LangGraph; earlier CrewAI coverage from 2026-06-07.
Linked by a graph relationship (LangGraph benchmarked against CrewAI); both cover LangGraph, LLM; earlier LangGraph coverage from 2026-03-15.
Linked by a graph relationship (LangGraph benchmarked against CrewAI); both cover CrewAI, LLM; earlier CrewAI coverage from 2026-02-27.
LangGraph uses OpenAI / Shared entity: LLM / Same source domain / Earlier coverage / Tension
Linked by a graph relationship (LangGraph uses OpenAI); both cover LLM; reported by the same outlet (arxiv.org).
LangGraph benchmarked against CrewAI / Shared entities / Earlier coverage
Linked by a graph relationship (LangGraph benchmarked against CrewAI); both cover CrewAI, LangGraph; earlier CrewAI coverage from 2026-03-01.
CrewAI supports MCP / Shared entity: LLM / Earlier coverage / Tension
Linked by a graph relationship (CrewAI supports MCP); both cover LLM; earlier LLM coverage from 2026-07-27.