Three tools with partial schema coverage and inconsistent description quality. Tool 1 (system_prompt_for_agent) has minimal description (67 chars, below baseline 194). Tool 2 (get_bedrock_logs) has good description (180 chars) and complete Pydantic schema with constraints (ge=1, le=90). Tool 3 (get_endpoint_metrics) has adequate description (145 chars) but schema visible only in signature, not in registration code. All tools are READ_ONLY and well-named with action verbs. No error handling guidance, no output schema documentation, no parameter descriptions in tool registration (only in Pydantic models). Naming follows verb_noun pattern (get_*, system_prompt_*) but descriptions lack WHEN/WHY context for LLM selection.

• Expand tool descriptions to 100-200 chars. Add WHEN/WHY context: 'Call this to generate a specialized system prompt for an AWS cost analysis agent. Use before initializing the agent to set cost-focused reasoning context.'
• Document output schemas explicitly. For get_bedrock_logs, specify: 'Returns DataFrame with columns: timestamp (ISO 8601), region (str), modelId (str), userId (str), inputTokens (int), completionTokens (int), totalTokens (int).'
• Add error handling guidance to descriptions. E.g., 'If logs are unavailable, check that the log group exists and the IAM role has logs:FilterLogEvents permission. Retry with a shorter time range.'
• Expose Pydantic model descriptions in FastMCP tool registration. Use @mcp.tool decorator with full parameter descriptions, not just names.
• Validate cross-account role assumption and return actionable errors: 'Failed to assume role arn:aws:iam::ACCOUNT:role/ROLE_NAME. Verify the role exists and the current identity has sts:AssumeRole permission.'