Shopping assistant agent powered by Gemini 3 Pro on Google ADK, reaching MongoDB Atlas through the official MongoDB MCP server and Anthropic's Fetch MCP server for public web data.
Five tools with basic schemas and descriptions, but significant gaps in parameter documentation, error handling, and LLM-optimized guidance. Tool names lack action verbs (find, aggregate, fetch are weak). Descriptions exist but are generic and under-specify constraints, formats, and recovery paths. No parameter descriptions for critical fields like 'filter' and 'pipeline', LLMs cannot infer MongoDB query syntax. No output schemas documented. No error handling guidance. Security concerns: write tools (insert_one, update_one) lack permission gates and audit trail documentation. Composition is reasonable (separate read/write tools) but parameter naming is inconsistent (collection vs. url).
Run MongoDB aggregation pipeline on a collection. Used to analyze purchase patterns and identify user preferences.
Retrieve and convert a URL to Markdown. Used to fetch public product specifications, competitor prices, and review summaries from the web.
Query MongoDB collection with filter, projection, sort, and limit. Used to retrieve documents from kajota database collections (users, products, purchases, wishlist).
Insert a single document into a MongoDB collection.
Update a single document in a MongoDB collection.
Tool names lack action verbs. 'find' and 'aggregate' are weak; should be 'query_collection' or 'search_documents'. 'fetch' is vague, does it download, parse, or convert? Rename to 'fetch_and_convert_url' or split concerns.
Critical parameters lack descriptions. 'filter' (MongoDB query object), 'pipeline' (aggregation stages), and 'document' (schema) have no guidance on format, constraints, or examples. LLMs cannot infer MongoDB syntax, they will hallucinate invalid queries.
No output schemas documented. Callers do not know what fields to expect from find/aggregate results. This breaks tool chaining, downstream tools cannot plan what data to extract.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 39 | 2026-07-28+ | v2 |
Write tools (insert_one, update_one) lack permission gates, audit trail documentation, and error recovery guidance. No mention of idempotency, dry-run, or confirmation steps. Agents could accidentally corrupt data.
No error handling guidance. What happens if a collection does not exist? If a filter is invalid? If MongoDB is unreachable? Descriptions do not explain recovery paths or error categories (retryable vs. fatal).