Resilience test: how six agent frameworks recover from eight broken tool calls

The agentic-arena project fed eight scripted tool-call faults to six agent frameworks. pydantic_ai, microsoft_af and smolagents recovered in all 8 cases, langgraph and openai_agents in 7, and google_adk in 6. smolagents burns up to 6 LLM calls on four faults versus 2 elsewhere.

• Eight faults: malformed JSON, unknown tool, missing argument, null instead of object
• pydantic_ai, microsoft_af and smolagents scored 8/8; langgraph and openai_agents 7/8
• google_adk scored 6/8 with uncaught exceptions on res-01 and res-02
• smolagents spends 6 LLM calls on four faults — about 2.7x the prompt tokens