[sundargthb]

Adds two small pieces for working with agents that already run on AgentCore Runtime. Both are framework agnostic.

EvaluationClient now allowlists the batch evaluation APIs (start, get, list, stop, delete) and adds start_batch_evaluation_and_wait, which scores sessions already in CloudWatch without invoking the agent. Scores come back under evaluationResults and Builtin.FailureAnalysis insights under insightResults. wait_until now accepts a set of target statuses so a run that ends as COMPLETED_WITH_ERRORS returns instead of timing out.

RecommendationClient passes through the four recommendation APIs (allowlisted, accepts snake_case kwargs) and adds start_recommendation_and_wait. It follows the same pattern as DatasetClient.

CloudWatchDataSourceConfig.to_data_source_config omits sessionIds when the list is empty, since an empty list filters to nothing.

batch = EvaluationClient(region_name="us-west-2").start_batch_evaluation_and_wait(
    wait_config=WaitConfig(max_wait=1800),
    batch_evaluation_name="nightly",
    evaluators=[{"evaluatorId": "Builtin.GoalSuccessRate"}, {"evaluatorId": "Builtin.FailureAnalysis"}],
    data_source_config={"cloudWatchLogs": {...}},
)
rec = RecommendationClient(region_name="us-west-2").start_recommendation_and_wait(
    wait_config=WaitConfig(max_wait=1800),
    name="promptrec",
    type="SYSTEM_PROMPT_RECOMMENDATION",
    recommendation_config={"systemPromptRecommendationConfig": {
        "systemPrompt": {"text": prompt},
        "agentTraces": {"batchEvaluation": {"batchEvaluationArn": batch["batchEvaluationArn"]}},
        "evaluationConfig": {"evaluators": [{"evaluatorArn": goal_success_arn}]},
    }},
)

Testing: unit tests added. The full suite passes (3621 passed). Tested live in us-west-2: GoalSuccessRate and FailureAnalysis ran over 2 existing runtime sessions in 102 s and returned 2 insight results, and a recommendation from a batch evaluation completed in 123 s.

[github-actions]

✅ No Breaking Changes Detected

No public API breaking changes found in this PR.

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[jariy17]

Two comments. RecommendationClient looks good. The one I care about is run_session_evaluation: I think it belongs on EvaluationClient with the batch APIs added to its allowlist, rather than on BatchEvaluationRunner, which exists for dataset runs. The other is a minor ordering nit.

[sundargthb]

Thanks, moved it. EvaluationClient now allowlists the five batch APIs and has start_batch_evaluation_and_wait, and the runner is back to main, so the ordering nit goes away with it. wait_until now also accepts a set of target statuses, since a batch evaluation can finish as COMPLETED or COMPLETED_WITH_ERRORS. The sessionIds fix stays. Verified live on two existing runtime sessions with GoalSuccessRate and FailureAnalysis.

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)