Paper Tests Five Agent Frameworks: Severe Flaws in Crash Recovery Semantics
Sajjad Khan · hf · 2026-08-06
This paper introduces the RESUME CONTRACT, a machine-checkable contract for examining the interrupt and resume semantics of agent workflow persistence layers. It defines six core properties (such as effect exactly-once, fork determinism, and checkpoint validity) and verifies them exhaustively using a TLA+ model.
The authors conducted deterministic tests on five widely deployed agent frameworks, revealing that they all answer recovery differently and contain severe flaws:
- LangGraph 1.2.9: Re-executes durably recorded work after a real crash, violating exactly-once semantics.
- CrewAI 1.15.2: Re-executes completed effect-bearing methods against its own written claims.
- pydantic-graph 1.x: Cannot resume at all after a mid-node crash.
Furthermore, concurrent recovery was found to be highly problematic, with multiple processes resuming a single parked interrupt causing effects to fire multiple times. To address this, the authors introduced REMIT, a verified reference sequencer that successfully repairs the identified failures.
More from coding & agent
- YC-Backed Axelrod Labs Launches: AI Agent Workforce to Automate Hotel Operations — ycombinator · 2026-08-06
- AI Agents Extracting Employee Knowledge Face Incentive Misalignment — manosaie · 2026-08-06
- Cloudflare Unveils Browser in Workers: 4x Less Memory for Millions of Concurrent Agents — dinasaur_404 · 2026-08-06
- AI Reshapes Dev Division: Frontend Roles to Merge with Product and Design — vista8 · 2026-08-06
- Stop Re-asking: A Practical Workflow for Organizing AI Coding Chats — Ok_Negotiation_2587 · 2026-08-06
- Cloudflare Kitesurf MCP Server in Action: Giving Agents a Browser — dinasaur_404 · 2026-08-06