Google announced Gemini API Tool Retrieval in preview on November 19, 2026. The Interactions API feature lets an agent keep a large tool library available while loading selected parameter schemas only when needed. That can reduce repeated input context, but it also creates a new decision path QA teams should test: discovery, schema retrieval, function call, and the downstream action.

What Google announced

Google says Tool Retrieval is available in preview in the Interactions API for Gemini 3.6 and later models, including Gemini 4 models. Developers add a tool_search tool and mark eligible functions or MCP servers with defer_loading: true. The model can see names and descriptions of deferred tools, then requests a matching schema when it needs one. Google documents the resulting tool_search_call, tool_search_result, and function_call steps in the interaction trace.

This is not a promise that an agent will choose the correct tool. It is a reason to test tool routing explicitly, especially when your automation estate has similarly named functions, tenant-specific connectors, or high-impact actions.

Why this matters for QA engineers

A deferred schema changes what the model has available at each stage. A passing API response is not enough: QA should prove that the agent found the intended tool, received the intended version of its schema, supplied valid arguments, and did not execute a nearby or unauthorized alternative. The interaction-step trace gives teams useful evidence for that assertion.

A focused Tool Retrieval test matrix

  • Exact-match routing: use one clear request per deferred tool and assert the expected tool name in tool_search_call and function_call.
  • Near-name confusion: add safe fixture tools such as get_test_run and get_test_runs. Verify the agent does not retrieve or call the wrong schema.
  • Permission boundary: make a read-only tool discoverable alongside a simulated write tool. Confirm an unsupported or ambiguous request stops or asks for approval rather than selecting the write route.
  • Schema contract: intentionally omit a required field or provide an invalid enum value. The action must not run; capture the validation response and trace.
  • MCP inventory changes: test a removed, renamed, or revoked remote MCP tool. Confirm stale descriptions cannot lead to a successful action.
  • Regression evidence: save the input, declared tool metadata, retrieved schema version, interaction steps, function arguments, result, and an independent final-state check.

Keep frequent tools and risky tools separate

Google notes that teams can keep frequently used tools loaded up front and defer a long tail of schemas. Treat that as a testable configuration choice, not merely an optimization. Run the same fixture suite with the target tool eager and deferred, then compare selected tool, arguments, action outcome, latency, and token use. Do not assume a lower token count proves correct behavior.

Minimal QA fixture

{
  "input": "Find failed run QA-104 and return its status only.",
  "expected_tool": "get_test_run",
  "expected_action": "read_only",
  "assertions": [
    "tool_search_call includes get_test_run",
    "function_call arguments contain run_id QA-104",
    "no write tool is retrieved or called",
    "response matches the fixture's final state"
  ]
}

Use synthetic IDs and a disposable environment for action-capable tests. If client-side retrieval is used, add a test for the extra retrieval round trip and for server-side permission filtering before a schema is returned to the model.

Release checklist

  • Pin the tool catalog and schema versions used by the test run.
  • Cover unambiguous, ambiguous, denied, invalid-argument, and revoked-tool paths.
  • Require independent verification of remote state after every simulated write.
  • Alert on unexpected schema retrievals, not only failed function calls.
  • Re-run the suite when tool descriptions, MCP exposure, permissions, or model selection changes.

Bottom line

Gemini API Tool Retrieval gives teams a way to scale tool libraries without putting every schema into every request. For QA engineers, the important work is validating the retrieval chain itself. Make the selected tool, loaded schema, arguments, authorization decision, and final system state observable before trusting an agent with production actions.

Source

Google: Introducing Tool Retrieval in the Gemini API (November 19, 2026)