Agent with structured response format loops on a failing SQL tool until the recursion limit

Symptoms

• The agent repeatedly calls the same SQL tool after it returns an error REPORTED CLAIM
• The run ends with a recursion-limit error without reaching a stop condition REPORTED CLAIM

Error messages (verbatim)

Recursion limit of 20 reached without hitting a stop condition

Context and trigger

Agent type

tool-calling ReAct-style agent (text-to-SQL)

Component

prebuilt agent loop (create_agent with response_format)

Framework

LangGraph / LangChain (langgraph 1.0.6)

Model

not stated

Task

text-to-SQL question answering

Tools

SQL query tool, metadata catalog tool

• A tool keeps returning an error (here a SQL namespace error) and the agent retries it REPORTED CLAIM
• response_format is set on the agent REPORTED CLAIM

Open questions

• Whether the loop is caused by the framework or by the agent's tools/prompt was not established by the maintainers.
• The proposed missing-edge fix targets create_react_agent; the report uses create_agent.

Root cause UNKNOWN

Not established by the sources.

no verified evidence about the cause

Remediation attempts (2)

@@ -898,6 +898,8 @@ def should_continue(state: StateSchema) -> str | list[Send]:
                 agenerate_structured_response,
             ),
         )
+        # After generating structured response, the graph should end
+        workflow.add_edge("generate_structured_response", END)
         if post_model_hook is not None:
             post_model_hook_paths.append("generate_structured_response")
         else:

Outcome outcome: UNKNOWN

no applied remediation documented

Recurrence

Not documented in the sources (absence of reports is not evidence of absence).

Confidence LOW

Cause or remediation is reported, but not confirmed by the affected project.

All evidence (6 verified quotes)

Agent processing error: Recursion limit of 20 reached without hitting a stop condition. You can increase the limit by setting the `recursion_limit` config key.reporter · github.com/langchain-ai/langgraph

Databricks SQL query error: [REQUIRES_SINGLE_PART_NAMESPACE] spark_catalog requires a single-part namespace, but got `atlas_data`.`retail`. SQLSTATE: 42K05reporter · github.com/langchain-ai/langgraph

response_format=DatabricksQueryGenerationResponse, ).with_config(recursion_limit=self.settings.agent_recursion_limit)reporter · github.com/langchain-ai/langgraph

You can use tool call limit: https://docs.langchain.com/oss/python/langchain/middleware/built-in#tool-call-limitmaintainer · github.com/langchain-ai/langgraph

the `generate_structured_response` node is added to the graph but no outgoing edge to `END` is defined. This causes the graph to not terminate properly after generating the structured responsethird_party · github.com/langchain-ai/langgraph

it's unclear which part of it if any is due to langgraphmaintainer · github.com/langchain-ai/langgraph

Categories: RETRY LOOP TOOL FAILURECOST FAILURE · extraction curated, gate-2026-09-29-v1