View a markdown version of this page

Generate a recommendation from insights findings - Amazon Bedrock AgentCore
Start a system prompt recommendationRetrieve the recommended promptAlternative trace sources
Services or capabilities described in Amazon Web Services documentation might vary by Region. To see the differences applicable to the China Regions, see Getting Started with Amazon Web Services in China (PDF).

Generate a recommendation from insights findings

After reviewing your insights results, you can use the same agent traces to generate an improved system prompt. Point StartRecommendation at the completed batch evaluation that produced your insights findings.

Note

Recommendation generation is currently supported only for Builtin.Insight.FailureAnalysis insights, and only system prompt recommendations are provided today.

Start a system prompt recommendation

Provide your current system prompt and reference the batch evaluation that contains the traces:

Example
AgentCore CLI

Generate a recommendation from a completed insights run:

agentcore run recommendation --from-insights <id> --type system-prompt --bundle-name my_bundle --json

Or reference a batch evaluation ARN directly:

agentcore run recommendation --batch-evaluation-arn <arn> --type system-prompt --bundle-name my_bundle --json
Interactive

  1. Run agentcore to open the TUI, then select run and choose Recommendation:

    Run menu: select Recommendation

    The wizard guides you through selecting a trace source (insights run or batch evaluation ARN), recommendation type, and config bundle.

Amazon SDK (boto3)
import boto3
import uuid

client = boto3.client("bedrock-agentcore", region_name="us-west-2")

response = client.start_recommendation(
    name=f"fix-rate-limiting-{uuid.uuid4().hex[:8]}",
    type="SYSTEM_PROMPT_RECOMMENDATION",
    recommendationConfig={
        "systemPromptRecommendationConfig": {
            "systemPrompt": {
                "text": "You are a helpful customer support assistant. Help users with their orders and returns."
            },
            "agentTraces": {
                "batchEvaluation": {
                    "batchEvaluationArn": "arn:aws:bedrock-agentcore:us-west-2:123456789012:batch-evaluate/insights-run-ABC1234567"
                }
            },
        }
    },
    clientToken=str(uuid.uuid4()),
)

recommendation_id = response["recommendationId"]
print(f"Started recommendation: {recommendation_id}")

Retrieve the recommended prompt

Poll until complete, then retrieve the result:

import time

while True:
    rec = client.get_recommendation(recommendationId=recommendation_id)
    if rec["status"] in ("COMPLETED", "FAILED"):
        break
    time.sleep(15)

if rec["status"] == "COMPLETED":
    result = rec["recommendationResult"]["systemPromptRecommendationResult"]
    print(f"Recommended prompt:\n{result['recommendedSystemPrompt']}")
    print(f"\nExplanation:\n{result['explanation']}")

Alternative trace sources

Instead of referencing a batch evaluation, you can also provide traces via CloudWatch Logs:

"agentTraces": {
    "cloudwatchLogs": {
        "logGroupArns": [
            "arn:aws:logs:us-west-2:123456789012:log-group:/aws/bedrock-agentcore/runtimes/MyAgent-abc123-DEFAULT"
        ],
        "serviceNames": ["MyAgent.DEFAULT"],
        "startTime": "2026-05-27T00:00:00Z",
        "endTime": "2026-06-03T00:00:00Z",
    }
}