diff --git a/openhands-sdk/openhands/sdk/agent/base.py b/openhands-sdk/openhands/sdk/agent/base.py
index dbd580e2d1..e97da7fb4a 100644
--- a/openhands-sdk/openhands/sdk/agent/base.py
+++ b/openhands-sdk/openhands/sdk/agent/base.py
@@ -304,6 +304,47 @@ def _decrypt_mcp_config(cls, data: Any, info: ValidationInfo) -> Any:
 
         return data
 
+    @model_validator(mode="after")
+    def _validate_unique_llm_usage_ids(self) -> AgentBase:
+        """Validate that LLMs with the same usage_id have the same configuration.
+
+        When multiple LLM references share a usage_id, they must have identical
+        settings. This commonly happens when:
+        1. An LLM object is intentionally shared (e.g., agent and condenser use
+           the same LLM) - this is fine, they're the same object
+        2. An agent is deserialized from JSON, creating separate LLM objects from
+           what was originally a shared reference - this is fine if configs match
+        3. Different LLMs are created with the same usage_id but different settings
+           (e.g., different models) - this is an error
+
+        Only case 3 raises an error, as silently using one LLM config while
+        ignoring another would be confusing.
+        """
+        usage_id_to_llms: dict[str, list[LLM]] = {}
+        for llm in self.get_all_llms():
+            usage_id_to_llms.setdefault(llm.usage_id, []).append(llm)
+
+        # Check for conflicting LLM configurations with the same usage_id
+        for usage_id, llms in usage_id_to_llms.items():
+            if len(llms) <= 1:
+                continue
+
+            # Compare LLM configs (excluding usage_id itself)
+            # Use model_dump with redaction to compare functional equivalence
+            first_config = llms[0].model_dump(exclude={"usage_id"})
+            for other_llm in llms[1:]:
+                other_config = other_llm.model_dump(exclude={"usage_id"})
+                if first_config != other_config:
+                    raise ValueError(
+                        f"Multiple LLMs share usage_id='{usage_id}' but have "
+                        f"different configurations. This would cause one LLM's "
+                        f"settings to be silently ignored. Set distinct usage_id "
+                        f"values for LLMs with different configs, e.g.: "
+                        f"LLM(model='gpt-4o', usage_id='agent'), "
+                        f"LLM(model='gpt-4o-mini', usage_id='condenser')"
+                    )
+        return self
+
     @model_serializer(mode="wrap")
     def _serialize_with_mcp_handling(self, handler, info: SerializationInfo):
         """Serialize the agent, handling mcp_config encryption/redaction.
diff --git a/openhands-sdk/openhands/sdk/conversation/impl/local_conversation.py b/openhands-sdk/openhands/sdk/conversation/impl/local_conversation.py
index a6bb1ec60e..948160847b 100644
--- a/openhands-sdk/openhands/sdk/conversation/impl/local_conversation.py
+++ b/openhands-sdk/openhands/sdk/conversation/impl/local_conversation.py
@@ -633,8 +633,18 @@ def _ensure_agent_ready(self) -> None:
             self.llm_registry.subscribe(self._state.stats.register_llm)
             registered = set(self.llm_registry.list_usage_ids())
             for llm in list(self.agent.get_all_llms()):
-                if llm.usage_id not in registered:
-                    self.llm_registry.add(llm)
+                if llm.usage_id in registered:
+                    # Skip duplicates but warn - this shouldn't happen if agent
+                    # was validated properly, but can occur with deserialized
+                    # agents that have the same usage_id on multiple LLMs
+                    logger.warning(
+                        f"Skipping duplicate LLM registration for usage_id "
+                        f"'{llm.usage_id}' (model={llm.model}). Consider using "
+                        f"distinct usage_id values for each LLM."
+                    )
+                    continue
+                self.llm_registry.add(llm)
+                registered.add(llm.usage_id)
 
             self._agent_ready = True
 
diff --git a/tests/sdk/llm/test_llm_registry.py b/tests/sdk/llm/test_llm_registry.py
index 9933a835f9..1f04e1c110 100644
--- a/tests/sdk/llm/test_llm_registry.py
+++ b/tests/sdk/llm/test_llm_registry.py
@@ -306,3 +306,160 @@ def test_llm_registry_does_not_reset_metrics_for_independent_llms():
     # llm2 should have its own independent metrics
     assert llm2.metrics is not llm1.metrics
     assert llm2.metrics.accumulated_cost == 0.0
+
+
+def test_agent_rejects_conflicting_usage_id_llms():
+    """Test agent validation rejects LLMs with same usage_id but different configs.
+
+    When an agent has multiple LLMs with the same usage_id but different
+    configurations (e.g., different models), validation should fail with a clear
+    error message guiding the user to set distinct usage_id values.
+
+    This catches the dangerous case where silently using one LLM config while
+    ignoring another would lead to unexpected behavior.
+    """
+    import pytest
+    from pydantic import SecretStr
+
+    from openhands.sdk.agent import Agent
+    from openhands.sdk.context.condenser import LLMSummarizingCondenser
+
+    # Create two separate LLM objects with the same usage_id but different models
+    agent_llm = LLM(
+        model="gpt-4o",
+        api_key=SecretStr("test-key"),
+        # usage_id defaults to "default"
+    )
+    condenser_llm = LLM(
+        model="gpt-4o-mini",  # Different model!
+        api_key=SecretStr("test-key"),
+        # usage_id defaults to "default" - same as agent_llm
+    )
+
+    # Verify they are different objects with the same usage_id
+    assert id(agent_llm) != id(condenser_llm)
+    assert agent_llm.usage_id == condenser_llm.usage_id == "default"
+
+    # Create condenser with its own LLM
+    condenser = LLMSummarizingCondenser(llm=condenser_llm, max_size=100)
+
+    # Agent construction should fail with a clear error message
+    with pytest.raises(ValueError) as exc_info:
+        Agent(llm=agent_llm, condenser=condenser, tools=[])
+
+    error_msg = str(exc_info.value)
+    assert "usage_id='default'" in error_msg
+    assert "different configurations" in error_msg
+
+
+def test_agent_accepts_shared_llm_with_same_usage_id():
+    """Test agent accepts when the same LLM is shared (e.g., with condenser).
+
+    When an agent shares the same LLM object between itself and its condenser,
+    this should work fine since they're the same object.
+    """
+    from pydantic import SecretStr
+
+    from openhands.sdk.agent import Agent
+    from openhands.sdk.context.condenser import LLMSummarizingCondenser
+
+    # Create a single LLM object shared between agent and condenser
+    shared_llm = LLM(
+        model="gpt-4o",
+        api_key=SecretStr("test-key"),
+    )
+
+    # Create condenser sharing the agent's LLM
+    condenser = LLMSummarizingCondenser(llm=shared_llm, max_size=100)
+
+    # Agent construction should succeed
+    agent = Agent(llm=shared_llm, condenser=condenser, tools=[])
+
+    # Verify only one LLM is discoverable (deduped by object identity)
+    llms = list(agent.get_all_llms())
+    assert len(llms) == 1
+    assert llms[0] is shared_llm
+
+
+def test_agent_accepts_duplicate_llms_with_identical_config():
+    """Test agent accepts multiple LLMs with same usage_id if configs match.
+
+    When an agent is deserialized from JSON, Pydantic creates separate LLM
+    objects even if they were originally the same object. As long as their
+    configurations match, this should not raise an error.
+    """
+    from pydantic import SecretStr
+
+    from openhands.sdk.agent import Agent
+    from openhands.sdk.context.condenser import LLMSummarizingCondenser
+
+    # Simulate what happens during deserialization: two separate objects
+    # with identical configs
+    agent_llm = LLM(
+        model="gpt-4o",
+        api_key=SecretStr("test-key"),
+        usage_id="shared",
+    )
+    condenser_llm = LLM(
+        model="gpt-4o",  # Same model
+        api_key=SecretStr("test-key"),  # Same key
+        usage_id="shared",  # Same usage_id
+    )
+
+    # Verify they are different objects
+    assert id(agent_llm) != id(condenser_llm)
+
+    # Create condenser with its own LLM
+    condenser = LLMSummarizingCondenser(llm=condenser_llm, max_size=100)
+
+    # Agent construction should succeed since configs match
+    agent = Agent(llm=agent_llm, condenser=condenser, tools=[])
+
+    # Verify both LLMs are discoverable (not deduped since different objects)
+    llms = list(agent.get_all_llms())
+    assert len(llms) == 2
+
+
+def test_agent_accepts_distinct_usage_id_llms():
+    """Test agent accepts LLMs with distinct usage_ids."""
+    import tempfile
+
+    from pydantic import SecretStr
+
+    from openhands.sdk.agent import Agent
+    from openhands.sdk.context.condenser import LLMSummarizingCondenser
+    from openhands.sdk.conversation import Conversation
+
+    # Create LLMs with distinct usage_ids
+    agent_llm = LLM(
+        model="gpt-4o",
+        api_key=SecretStr("test-key"),
+        usage_id="agent",
+    )
+    condenser_llm = LLM(
+        model="gpt-4o-mini",
+        api_key=SecretStr("test-key"),
+        usage_id="condenser",
+    )
+
+    # Create condenser with its own LLM
+    condenser = LLMSummarizingCondenser(llm=condenser_llm, max_size=100)
+
+    # Agent construction should succeed
+    agent = Agent(llm=agent_llm, condenser=condenser, tools=[])
+
+    # Verify both LLMs are discoverable
+    llms = list(agent.get_all_llms())
+    assert len(llms) == 2
+    usage_ids = {llm.usage_id for llm in llms}
+    assert usage_ids == {"agent", "condenser"}
+
+    with tempfile.TemporaryDirectory() as tmpdir:
+        # Create conversation - this should work fine
+        convo = Conversation(agent=agent, workspace=tmpdir, persistence_dir=tmpdir)
+        convo.send_message("Hello")
+
+        # Verify both LLMs were registered
+        assert "agent" in convo.llm_registry.list_usage_ids()
+        assert "condenser" in convo.llm_registry.list_usage_ids()
+        assert len(convo.llm_registry.list_usage_ids()) == 2