feat: Add interface for agent optimizers

Details:
* The Sampler allows ADK agent optimizers to request evals and receive detailed eval results that can be used to guide agent optimization.
* This interface allows developers to run custom evaluations if needed for their agent. An implementation to interface with the inbuilt LocalEvalService shall be published shortly in a follow-up PR.
* The AgentOptimizer interface describes the general framework for an ADK agent optimizer that uses the Sampler interface to optimize an ADK agent.
* data_types.py contains generic types and base classes to allow intercommunication between the AgentOptimizer, Sampler, and developer code.

Co-authored-by: Keyur Joshi <keyurj@google.com>
PiperOrigin-RevId: 861849691

Author: KPJoshi

src/google/adk/optimization/__init__.py

@@ -0,0 +1,13 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

src/google/adk/optimization/agent_optimizer.py

@@ -0,0 +1,49 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from abc import ABC
+from abc import abstractmethod
+from typing import Generic
+
+from ..agents.llm_agent import Agent
+from .data_types import AgentWithScores
+from .data_types import OptimizerResult
+from .data_types import SamplingResult
+from .sampler import Sampler
+
+
+class AgentOptimizer(ABC, Generic[SamplingResult, AgentWithScores]):
+  """Base class for agent optimizers."""
+
+  @abstractmethod
+  async def optimize(
+      self,
+      initial_agent: Agent,
+      sampler: Sampler[SamplingResult],
+  ) -> OptimizerResult[AgentWithScores]:
+    """Runs the optimizer.
+
+    Args:
+      initial_agent: The initial agent to be optimized.
+      sampler: The interface used to get training and validation example UIDs,
+        request agent evaluations, and get useful data for optimizing the agent.
+
+    Returns:
+      The final result of the optimization process, containing the optimized
+      agent instances along with their corresponding scores on the validation
+      examples and any optimization metadata.
+    """
+    ...

src/google/adk/optimization/data_types.py

@@ -0,0 +1,90 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from typing import Any
+from typing import Generic
+from typing import Optional
+from typing import TypeVar
+
+from google.adk.agents.llm_agent import Agent
+from pydantic import BaseModel
+from pydantic import Field
+
+
+class BaseSamplingResult(BaseModel):
+  """Base class for evaluation results of the candidate agent on the batch of examples."""
+
+  scores: dict[str, float] = Field(
+      required=True,
+      description=(
+          "A map from example UID to the agent's overall score on that example."
+          " (higher is better)."
+      ),
+  )
+
+
+# SamplingResult: the per-component evaluation results for a batch of examples.
+# Should at least include per-example scores and may also contain other data
+# required for optimizing the agent (e.g., outputs, trajectories, and metrics).
+SamplingResult = TypeVar("SamplingResult", bound=BaseSamplingResult)
+
+
+class BaseAgentWithScores(BaseModel):
+  """An optimized agent with its scores.
+
+  Optimizers may use the overall_score field and can return custom metrics by
+  sub-classing this class.
+  """
+
+  optimized_agent: Agent = Field(
+      required=True,
+      description="The optimized agent.",
+  )
+
+  overall_score: Optional[float] = Field(
+      default=None,
+      description="The overall score of the optimized agent.",
+  )
+
+
+AgentWithScores = TypeVar("AgentWithScores", bound=BaseAgentWithScores)
+
+
+class OptimizerResult(BaseModel, Generic[AgentWithScores]):
+  """Base class for optimizer final results."""
+
+  optimized_agents: list[AgentWithScores] = Field(
+      required=True,
+      description=(
+          "A list of optimized agents which cannot be considered strictly"
+          " better than one another (see"
+          " https://en.wikipedia.org/wiki/Pareto_front), along with scores."
+      ),
+  )
+
+
+class UnstructuredSamplingResult(BaseSamplingResult):
+  """Evaluation result providing per-example unstructured evaluation data."""
+
+  data: Optional[dict[str, dict[str, Any]]] = Field(
+      default=None,
+      description=(
+          "A map from example UID to JSON-serializable evaluation data useful"
+          " for agent optimization. Recommended contents include inputs,"
+          " trajectories, and metrics. Must be provided if requested by the"
+          " optimizer."
+      ),
+  )

src/google/adk/optimization/sampler.py

@@ -0,0 +1,70 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import annotations
+
+from abc import ABC
+from abc import abstractmethod
+from typing import Generic
+from typing import Literal
+from typing import Optional
+
+from ..agents.llm_agent import Agent
+from .data_types import SamplingResult
+
+
+class Sampler(ABC, Generic[SamplingResult]):
+  """Base class for agent optimizers to sample and score candidate agents.
+
+  The developer must implement this interface for their evaluation service to
+  work with the optimizer. The optimizer will call the sample_and_score method
+  to get evaluation results for the candidate agent on the batch of examples.
+  """
+
+  @abstractmethod
+  def get_train_example_ids(self) -> list[str]:
+    """Returns the UIDs of examples to use for training the agent."""
+    ...
+
+  @abstractmethod
+  def get_validation_example_ids(self) -> list[str]:
+    """Returns the UIDs of examples to use for validating the optimized agent."""
+    ...
+
+  @abstractmethod
+  async def sample_and_score(
+      self,
+      candidate: Agent,
+      example_set: Literal["train", "validation"] = "validation",
+      batch: Optional[list[str]] = None,
+      capture_full_eval_data: bool = False,
+  ) -> SamplingResult:
+    """Evaluates the candidate agent on the batch of examples.
+
+    Args:
+      candidate: The candidate agent to be evaluated.
+      example_set: The set of examples to evaluate the candidate agent on.
+        Possible values are "train" and "validation".
+      batch: List of UIDs of examples to evaluate the candidate agent on. If not
+        provided, all examples from the chosen set will be used.
+      capture_full_eval_data: If false, it is enough to only calculate the
+        scores for each example. If true, this method should also capture all
+        other data required for optimizing the agent (e.g., outputs,
+        trajectories, and tool calls).
+
+    Returns:
+      The evaluation results, containing the scores for each example and (if
+      requested) other data required for optimization.
+    """
+    ...