AI Tools & Technology

Gemini Agents Get a Tool Finder to Cut Through Software Clutter

British newspaper cartoon showing a Google Gemini AI robot searching through a chaotic collection of software tools with a magnifying glass while an overwhelmed office worker looks on.
Too many tools, too little clarity. Google's Gemini agents gain a way to identify the software they need without getting lost in an increasingly crowded digital toolbox.

An AI assistant that can use a handful of software tools is relatively easy to describe. Give it access to hundreds of functions and the picture becomes much more complicated. The system needs to know which tool is relevant, what information it requires and whether the action is permitted. Google has introduced Gemini Tool Retrieval in its API to reduce the burden of feeding every detailed tool description into every request.

In its 19 November announcement, Google says developers can keep a large tool catalogue available while loading the full parameter instructions only when a particular function is needed. The feature is in preview in the Gemini Interactions API and works with the stated range of recent Gemini models. It is a developer tool, not a new button that ordinary Gemini users will necessarily see.

Why Gemini Tool Retrieval can help agents

Imagine a company assistant connected to customer records, invoices, calendars, deliveries and support systems. To perform one task, it may need only a tool that finds an order and another that checks its status. Sending the complete technical instructions for every available function wastes space and creates more opportunities for confusion. The agent may have to choose among many functions with similar names and overlapping capabilities.

The issue is not just computing cost. A mistaken tool choice can change the outcome of a task. A function that drafts a refund is not the same as one that issues it; a search tool should not be confused with an action that edits records. Clear names and parameter definitions help, but an enormous list can still overwhelm a model’s immediate context.

Google’s approach is to make tools discoverable without forcing all their details into the initial conversation. The model can identify the function it needs, retrieve the relevant schema and then call it. The idea resembles a skilled worker using a well-organised reference manual instead of memorising every procedure at once.

What developers still have to specify

The new mechanism does not invent a reliable business process automatically. Someone still has to define functions accurately, specify what inputs they require and control which actions may be performed. A vaguely named tool or badly designed permission system remains dangerous even if the model retrieves its instructions efficiently.

Google’s Interactions API announcement describes the wider developer environment in which tools can be called and responses handled. Developers should verify the current tool-retrieval feature’s availability and syntax in the documentation for the Interactions API rather than copying an early example without checking updates. Preview features can change.

The architecture is particularly relevant when an AI agent can take action rather than merely return a written answer. A tool for reading a record should not have the permissions of a tool for deleting one. An agent should not acquire authority simply because it has discovered that a function exists. Discovery and permission are separate security decisions.

The difference between token savings and better judgement

Google says loading fewer full schemas can reduce input tokens and latency while helping accuracy. That claim is plausible from the design: unnecessary descriptions occupy space and cost money. But the exact saving depends on how many functions a typical task actually needs. If an agent repeatedly loads most of a small tool library, there may be little benefit.

A development team can test the feature by replaying representative tasks with and without on-demand loading. It should record successful completion, wrong-tool calls, time, total cost and how often the agent needs clarification. Comparing only token counts could hide a change that makes the system cheaper but less dependable.

LiveAIWire’s exploration of AI workflows that leave employees with extra work illustrates why a technical improvement must eventually benefit the person using the system. Saving a fraction of a second is valuable at scale, but it does not justify extra correction work or an unexpected transaction.

Why this could matter outside software teams

Ordinary users may never encounter the phrase Tool Retrieval, but they will notice whether an assistant can pick the correct application or function. A travel assistant that checks the wrong calendar, a purchasing assistant that misreads a budget and a finance agent that selects the wrong export each produce a recognisable problem. Better tool selection is one ingredient in making these systems more predictable.

It also affects how organisations decide to connect new services. An assistant with hundreds of available functions should not automatically get the broadest access to all of them. Each connection needs a business justification and a controlled approval path. An accurate function call is still an unauthorised call if the user did not grant permission.

LiveAIWire has previously looked at AI agents working across email and calendars. Those workflows are simple enough to explain but complex enough to expose the distinction between finding the right tool and using it responsibly.

Google’s announcement is a practical step in the plumbing of agent software. It will not attract the attention of a spectacular new chatbot model, but it addresses one of the reasons real-world assistants can be inefficient and unreliable. The test is whether the design makes useful actions more accurate without giving software more authority than it needs.

About the Author

Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.