Multi-agent system for security operations with detection-as-code rule tuning, threat hunting, alert triage, and incident response workflows using Google ADK and MCP tools
22 tools with mixed quality. Naming is generally verb-first and clear (get_*, create_*, run_*, load_*). Descriptions present for all tools but many are generic or lack actionable context. Schemas visible for most tools with proper JSON Schema structure, but several lack parameter descriptions or have incomplete type definitions. Error handling guidance is minimal. Output schemas are not documented. No tool annotations (readOnlyHint/destructiveHint). Several tools expose internal complexity (MCP tool execution, workflow invocation) without clear user-facing abstractions.
Creates a GitHub Pull Request using GitHub CLI (gh).
Executes an MCP security tool dynamically with the supplied arguments.
Gets the current time, formatted for use in filenames or timestamps.
Returns a summary of the execution metrics (time, tokens) collected so far. Use this tool to retrieve data for the 'Operational Artifacts' section of reports.
Retrieves the full JSON Schema, parameter definitions, and descriptions for an MCP tool. Use this tool to inspect required parameters, types, and schema details before executing an MCP tool via `execute_mcp_tool`.
Stages specified files and commits them with the given message.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint). LLMs cannot distinguish safe read-only tools from destructive operations. Git operations (commit, push, create_github_pr) and workflow runners should be marked destructive; read operations should be marked readOnly.
Output schemas not documented. Tools like execute_mcp_tool, run_*_workflow, and load_skill return complex results but no schema is provided. LLMs cannot plan downstream tool calls or extract required fields without knowing the response structure.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 68 | <=2025-11-25 | v2 |
Creates a new git branch in the specified repository.
Pushes a branch to the remote repository.
Lists available progressive disclosure skills, optionally filtered by category. Use this tool to discover available security skills and capabilities that you can load on-demand using `load_skill`.
Loads persona description and appends progressive disclosure skill catalog. Reads the persona markdown file and appends the catalog of available skills so the agent is aware of what skills it can dynamically load during execution.
Loads the full markdown instructions, procedures, and rubrics for a specified skill. Use this tool when you need complete step-by-step procedures, execution guidelines, or validation rubrics to perform a security investigation, threat hunt, alert triage, or incident response task.
Reads the content of a file.
Executes ADK Graph Workflow: advanced_threat_hunting_workflow.
Executes ADK Graph Workflow: alert_report_workflow.
Executes ADK Graph Workflow: apt_threat_hunt_workflow.
Executes ADK Graph Workflow: basic_ioc_enrichment_workflow.
Executes ADK Graph Workflow: case_report_workflow.
Executes ADK Graph Workflow: close_duplicate_cases_workflow.
Executes ADK Graph Workflow: cloud_vulnerability_triage_workflow.
Discovers available MCP security tools by keyword or server without loading full schemas. Use this tool to find relevant MCP tools for SIEM (Chronicle), SOAR (SecOps SOAR), or GTI (VirusTotal) before requesting their schema or executing them.
Validates YAML syntax of a detection rule file.
Writes a report to a file in the designated reports directory. Saves the given report content to a markdown file with automatic timestamp appending if not already present.
Generic descriptions lack actionable context. Tools like read_file_content, run_case_report_workflow, and validate_yaml_file have minimal descriptions (50-55 chars). LLMs need WHEN to use, WHAT it returns, and any prerequisites. Example: 'Reads the content of a file.' should be 'Reads and returns the full text content of a file. Use this to inspect configuration, logs, or detection rules before processing. Returns raw file content as a string.'
No error handling guidance. Tools do not document what errors can occur, whether they are retryable, or what the LLM should do next. Example: execute_mcp_tool could fail if the tool name is invalid, the arguments don't match the schema, or the MCP server is unreachable, but no guidance is provided.
Parameter descriptions missing or incomplete. Several tools have parameters without descriptions (e.g., execute_mcp_tool's 'arguments' param is described as 'Dictionary of arguments matching the tool's input schema (or JSON string)' but does not explain when to use dict vs string, or how to handle nested objects). LLMs cannot infer parameter meaning from names alone.