AI News/AI News

Gemini API adds Tool Retrieval to keep large agent tool sets lean

Google's preview feature loads tool details only when an agent needs them, which matters for teams wiring marketing agents to many tools.

Key points

  • Tool Retrieval is in preview in the Gemini Interactions API; developers add a tool_search tool and mark tools with defer_loading set to true.
  • Deferred tools still show name and description, while full parameter schemas are fetched by the API only when the model asks for one.
  • Supported on Gemini 3.6 and newer, which covers every Gemini 4 model; no separate fee, standard token rates apply.
  • Teams with their own registry or permission filters can run the search client-side, which costs one extra round trip per lookup.

Why it matters for marketers: Teams building agents with many connected tools can cut prompt bloat and keep tool choice reliable, without building or hosting a search index of their own.

Google has released Tool Retrieval in preview inside the Gemini Interactions API. Instead of sending every tool's full parameter schema with each request, an agent holds a library on standby and pulls in details only for the tool a task calls for. Google says this lowers input tokens and latency, and helps the model pick the right tool as the list grows.

Setup is light. Developers add a tool_search tool and flag the functions they want held back. The model still sees each deferred tool's name and description, so it knows what exists. When it needs one, it requests it by name, the API loads the schema, and the call happens within the same request. Google says there is no index to build or host. Remote MCP servers are covered too, so one flag can defer everything a server exposes.

For marketing teams, the practical effect is on agents connected to many systems, such as a CRM, an analytics platform, a CMS, an ad account and an email tool. Each added integration usually makes every prompt heavier. Deferring the long tail keeps costs and response times in check, while frequently used tools can stay loaded up front.

Pricing is unchanged: usage is billed at the model's standard token rates. Google notes that enabling the search adds a small fixed block of input tokens, and any schema the model loads is billed for the rest of the conversation. Savings depend on how many tools go unused, so Google advises measuring against your own traffic. Teams with custom registries or per-user permissions can run the search on their own side, trading one extra round trip per search for control.

Sources

This brief is our own summary. Read the full reporting at the source.

You read about AI automation. We run it.

DocFluence, built by Z Media, writes, designs, films and publishes for your brand every day, in your voice, from one login.