MCP server that delegates coding and research tasks to Google Antigravity sub-agents
Single tool 'delegate_to_antigravity' has a clear, descriptive name starting with a verb and a detailed description (180+ chars). Input schema is fully visible with 8 parameters, all typed and described. However, output schema is not documented, the tool returns 'the sub-agent's final report' but the structure of that report is never specified. Error handling is present (ToolError for validation) but lacks recovery guidance. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite the tool having WRITE risk. The tool is well-engineered internally but falls short of production-grade agent tool patterns.
Delegates a coding or research task to Google Antigravity, running the agy CLI as a subprocess and returning the sub-agent's final report.
Output schema not documented. Tool description says it returns 'the sub-agent's final report' but does not specify the structure (fields, types, example). LLMs cannot plan downstream actions or extract data without knowing what fields to expect.
No tool annotations despite WRITE risk. Tool modifies files and runs shell commands (when allow_edits=true, allow_shell=true) but lacks destructiveHint or idempotentHint annotations. Agents cannot reason about retry safety or side effects.
Error messages lack recovery guidance. Validation errors (e.g. 'timeout_seconds must be positive') tell the LLM what failed but not what to do next. Should include hints like 'Pass a value between 1 and 3600 seconds.'
Parameter descriptions are verbose (some 150+ chars). Baseline is 72 chars avg. Trim to essential info: 'Absolute path to the directory the sub-agent should work in' instead of repeating 'Pass an absolute path...' in validation logic.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 67 | <=2025-11-25 | v2 |
No idempotency guarantee documented. If a delegation times out and the agent retries, will the sub-agent re-run the same task or resume? Agents need to know if this tool is safe to retry.