Overview

• AWS credentials configured with permissions described in minimum_iam_policy.md
• The SDK environment has been verified (SDK version, region, execution role). If not done, activate the sdk-getting-started skill first.

Principles

1. Read-only: No mutations — only Describe, GetMetricData, and FilterLogEvents calls
2. Deterministic: No heuristics, no scoring, no classification
3. Graceful degradation: Each collection step is independent; failures in one do not block others
4. Agent interprets: The script collects facts; the agent provides interpretation and guidance
5. First-variant metrics only: Instance-level metrics (CPU, Memory, GPU) are collected for the first production variant only. For multi-variant endpoints, the agent should note this limitation when presenting results.

Trigger

Activate when the user:

• Reports endpoint issues, errors, or latency (inference-time problems)
• Asks to check endpoint health, status, or metrics
• Wants to debug inference failures or timeouts on a deployed endpoint
• Reports a deployment failure (endpoint creation failed)
• Asks about instance count, container logs, or resource utilization of an endpoint

Do NOT activate for

• Training job failures — use the finetuning skill instead. Training jobs and endpoints are separate SageMaker resources.
• Listing, creating, updating, or deleting endpoints — this skill diagnoses existing endpoints, not endpoint lifecycle management.
• Model deployment requests — use the model-deployment skill instead.
• Scaling or capacity changes — this skill collects diagnostics, it does not modify endpoints.

Requirements

• Endpoint name: The SageMaker endpoint to diagnose
• AWS region: The region where the endpoint is deployed

Workflow

Step 1: Collect inputs

For this step, you need the endpoint name and AWS region:

1. Check conversation history — the user may have already mentioned the endpoint name or region.
2. Silently read project files (e.g., deployment notebooks, config files, sdk-getting-started output) for the region or endpoint name.
3. Only if still unknown, ask the user for the missing values.

⏸ Wait for user response if any values are missing.

Step 2: Run diagnostics

Execute collect_diagnostics.py with the endpoint name and region. Do not create a notebook — run the script directly:

The script collects:

• Endpoint status via DescribeEndpoint
• CloudWatch metrics (invocations, errors, latency, utilization) for the last 5 minutes
• Container logs from the last 15 minutes (up to 100 events)

Step 3: Present results

Present the collected data to the user. For any issues found, reference the official AWS troubleshooting guide: https://docs.aws.amazon.com/sagemaker/latest/dg/deploy-model-troubleshoot.html (opens in a new tab)

References

• Minimum IAM Policy
• AWS Endpoint Troubleshooting Guide (opens in a new tab)
• CloudWatch Metrics for SageMaker (opens in a new tab)