How to Create a Diagnostic Agent with Gemini and OSQuery

Talk delivered at DevFest Scotland 2025.

Transcript

1. Scotland Daniela Petruzalek Developer Relations, Google Cloud How to Create a Diagnostic Agent with Gemini and OSQuery
2. About me… DevRel at Google UK Originally from Brazil Backend / Data Engineer Currently obsessed with AI Love Games, Anime and Cats =^_^=
3. Agenda Introduction Agent Development Kit OSQuery The Evolution of AIDA
4. Introduction
5. “Hello computer”
6. Agent Development Kit A framework for development and deployment of AI agents. Optimised for Gemini and Google products, but also model-agnostic and environment-agnostic google.github.io/adk-docs
7. OSQuery A monitoring, instrumentation and analysis tool for operating systems using the SQL language Available for Linux, macOS and Windows. www.osquery.io
8. Architecture v1
9. root agent OSQuery OS Dev UI (runner) user run_osquery Gemini 2.5 Flash
10. // root_agent is the entry point for an ADK agent root_agent = Agent( model="gemini-2.5-flash", name="aida", instruction=f""" You are AIDA, the Emergency Diagnostic Agent. - Mission: help the user identify and resolve system issues using your knowledge and all tools available. - Host OS: {platform.system().lower()} """, tools=[ run_osquery, ], )
11. def run_osquery(query: str) -> str: """Runs a query using osquery. Args: query: Query to run, e.g., 'select * from battery' Returns: the query result as a JSON string. """ result = subprocess.run( ["osqueryi", "--json", query], capture_output=True, text=True, timeout=60 ) output = result.stdout.strip() return output
12. Architecture v2
13. Schema discovery Response quality depends on context quality Idea: use Retrieval Augmented Generation (RAG) for table schemas github.com/osquery/osquery/tree/master/specs
14. root agent OSQuery OS Dev UI (runner) user run_osquery Gemini 2.5 Flash SQLite RAG discover_schema NEW!
15. from sqlite_rag import SQLiteRag def ingest(rag: SQLiteRag, file_path: str): with open(file_path, "r", encoding="utf-8") as f: content = f.read() rel_path = os.path.relpath(file_path, SPECS_DIR) rag.add_text(content, uri=rel_path) # in __main__ rag = SQLiteRag.create(DB_PATH, settings={"quantize_scan": True}) files_to_ingest = [...] # omitted for brevity for i, file_path in enumerate(files_to_ingest): ingest(rag, file_path) rag.quantize_vectors() rag.close()
16. schema_rag = SQLiteRag.create(SCHEMA_DB_PATH) def discover_schema(terms: str, platform: str, top_k: int): """ Queries the osquery schema documentation and returns all table candidates to explore the provided search terms. Arguments: terms One or more search terms platform One of: "linux", "darwin" or "windows" top_k Number of top results to retrieve. Returns: up to top_k related table schemas. """ terms += " " + platform return schema_rag.search(terms, top_k=top_k)
17. // root_agent is the entry point for an ADK agent root_agent = Agent( model="gemini-2.5-flash", name="aida", instruction=f""" ... - If a query returns an empty result, use discover_schema to certify that the query is correct """, tools=[ run_osquery, discover_schema ], )
18. Architecture v3
19. Query packs Instead of raw schema information, give the agent curated queries for common use cases github.com/osquery/osquery/tree/master/packs
20. root agent OSQuery OS Dev UI (runner) user run_osquery Gemini 2.5 Flash SQLite RAG discover_schema NEW! SQLite RAG search_query_library
21. def search_query_library(terms: str, platform: str, top_k: int): """Search the query library to find queries corresponding to the search terms. Arguments: terms One or more search terms platform One of: "linux", "darwin" or "windows" top_k Number of documents to return Returns: up to top_uk best queries based on the search terms """ terms += " " + platform return queries_rag.search(terms, top_k=top_k)
22. // root_agent is the entry point for an ADK agent root_agent = Agent( model="gemini-2.5-flash", name="aida", instruction=f""" ... - Always search the query library for useful queries """, tools=[ run_osquery, discover_schema, search_query_library ], )
23. Architecture v4
24. External knowledge Built-in tools: Google Search, code execution, Vertex AI RAG Engine, Vertex AI Search, Big Query, … Only one built-in tool is supported per agent But agents can call other agents: AgentTool
25. from google.adk.tools.google_search_tool import google_search from google.adk.tools.agent_tool import AgentTool search_agent = Agent( model=MODEL, name="search_agent", description="An agent specialised in searching the web", instruction=f""" Use the google_search tool to fulfill the request. Always return the complete information including examples. """, tools=[ google_search ], ) search_tool = AgentTool(search_agent)
26. // root_agent is the entry point for an ADK agent root_agent = Agent( model="gemini-2.5-flash", name="aida", instruction=f""" ... - If the osquery tool do not provide enough information, use the search_tool to expand your knowledge. """, tools=[ run_osquery, discover_schema, search_query_library, search_tool ], )
27. root agent OSQuery OS Dev UI (runner) user run_osquery Gemini 2.5 Flash SQLite RAG discover_schema NEW! SQLite RAG search_query_library Google Search search_tool
28. Architecture v5
29. Agent Patterns
30. Dev UI (runner) user Coordinator (root agent) OS Specialist OSQuery Specialist Google Search SQLite RAG (query library) SQLite RAG (schemas) OSQuery Pipeline Query Planner
31. os_agent = Agent(name="os_specialist", ...) osquery_agent = Agent(name="osquery_specialist", ...) pipeline = SequentialAgent( name="diagnostic_pipeline", sub_agents=[os_agent, osquery_agent] ) root_agent = Agent(name="aida", ..., instruction=f"""... - When the user makes a diagnostic request, delegate it to the diagnostic_pipeline """, sub_agents=[pipeline] )
32. The Client Interface
33. Frontend (runner) user Coordinator (root agent) OS Specialist OSQuery Specialist Google Search SQLite RAG (query library) SQLite RAG (schemas) OSQuery Pipeline Query Planner
34. from aida_v5.agent import root_agent ss = InMemorySessionService() runner = Runner(app_name="aida", agent=root_agent, session_service=ss) @app.post("/chat") async def chat_handler(request: Request): # ... boring parts here, e.g. session management ... async def stream_generator(): full_response = "" async for event in runner.run_async( user_id=user_id, session_id=session_id, new_message=user_query, ): # ... process event parts and build response... return StreamingResponse(stream_generator(), ...)
35. github.com/sito-sikino/avatar-ui-core
36. Final words This is just the beginning for AIDA: Better architecture: Loop, Parallel, … Application specific knowledge Memory Kubequery Run commands (danger zone!!!)
37. Scotland danicat.dev/posts/20251031-building-aida danicat.dev/posts/20251103-building-aida-part-2 Examples: github.com/danicat/devfest2025 Daniela Petruzalek Developer Relations, Google Cloud danicat.dev | @danicat83 Thank you!