fix: skip AgentState reset for MultiAgentBase executors in GraphNode

[giulio-leone]

Problem

GraphNode.reset_executor_state() unconditionally overwrites executor.state with AgentState(...), corrupting the GraphState dataclass on nested Graph/Swarm executors.

Two call sites are affected:

1. _execute_node (guarded by reset_on_revisit): triggers when a nested graph node is revisited in a cycle
2. deserialize_state (unguarded): iterates all nodes unconditionally when restoring a completed graph run

Root Cause

__post_init__ correctly guards against capturing MultiAgentBase state (hasattr(self.executor.state, "get")), so _initial_state stays as the default empty AgentState(). But reset_executor_state lacks this guard:

# Before (corrupts GraphState on MultiAgentBase executors)
if hasattr(self.executor, "state"):
    self.executor.state = AgentState(self._initial_state.get())

Fix

Add isinstance(self.executor, MultiAgentBase) guard to skip the state reset for multi-agent executors whose state is not an AgentState dict:

# After (preserves GraphState on MultiAgentBase executors)
if hasattr(self.executor, "state") and not isinstance(self.executor, MultiAgentBase):
    self.executor.state = AgentState(self._initial_state.get())

Test

Added test_reset_executor_state_preserves_multiagent_state_type that:

1. Creates a MultiAgentBase mock with a GraphState attribute
2. Calls reset_executor_state()
3. Asserts the state remains a GraphState instance (not overwritten to AgentState)

Fixes #1775

[Copilot]

Pull request overview

Fixes a state-type corruption bug in GraphNode.reset_executor_state() where nested multi-agent executors (e.g., Graph/Swarm) could have their dataclass state overwritten with an AgentState during node resets and session deserialization.

Changes:

• Skip resetting executor.state when the executor is a MultiAgentBase, preserving GraphState/SwarmState types.
• Add a regression test ensuring reset_executor_state() does not overwrite a multi-agent executor’s state type.
• Expand method docstring to clarify why multi-agent state resets are skipped.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/strands/multiagent/graph.py	Adds MultiAgentBase guard in GraphNode.reset_executor_state() to prevent clobbering nested executor state types.
tests/strands/multiagent/test_graph.py	Adds regression test for #1775 validating multi-agent state type is preserved across resets.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

NodeResult.result is defined as AgentResult | MultiAgentResult | Exception, but this test sets it to a plain string. This can make the test brittle (e.g., any future code path calling NodeResult.to_dict()/get_agent_results() would fail) and doesn’t reflect real usage. Consider constructing a minimal AgentResult (or MultiAgentResult) here instead of a string.

Suggested change

	node.result = NodeResult(
	result="test",
	agent_result = AgentResult(
	message={"role": "assistant", "content": [{"text": "test"}]},
	stop_reason="end_turn",
	state={},
	metrics=None,
	)
	node.result = NodeResult(
	result=agent_result,

[giulio-leone]

Fixed in 52810a0 — replaced plain string with proper AgentResult instance matching the type annotation.

[giulio-leone]

All CI checks pass. Ready for review.