What is TSO (tool selection optimization)?
TSO (Tool Selection Optimization) is an operational framework proposed by LS-Labs for checking why an AI agent, while carrying out a task, did or did not choose a particular tool, API or MCP server.
It splits that process into four stages: integration, exposure, selection and execution.
The four stages
| Stage | What happens |
|---|---|
| Integration | The provider connects the tool to a system the agent can use. In MCP, a tool definition includes a name, a description and a JSON Schema for its input parameters. Models choose tools from this metadata, not from the code behind them. |
| Exposure | With many tools, the system searches first and puts only some tool definitions into the model's context. With Anthropic's tool search, for example, tools marked for deferred loading are loaded only when Claude finds them through search. An unloaded tool cannot be chosen. |
| Selection | The model picks from the loaded candidates. The BiasBusters study found that semantic alignment between the user's query and a tool's metadata is the strongest driver of selection, and that small changes to a tool's description can significantly shift choices. |
| Execution | Once the model settles on a tool and its arguments, the client sends the call, for example with MCP's tools/call request. The right tool with a wrong argument still fails, so record selection and task completion separately. |
Check each stage to find where it breaks
When an agent does not use a tool, the tool may never have reached the candidate list, so the model never saw it. Or it made the list and the model picked a different tool. To cope with large numbers of tools, Anthropic, OpenAI and Microsoft Foundry all offer tool search: search first, then load the needed tool definitions into the model's context.
Suppose a restaurant's booking API is connected to an AI agent, a user asks to book a table, and the API is never called. If search never returned it as a candidate, the problem is exposure; if it lost to another booking tool, look at selection; if it was chosen but got the date wrong, the problem is execution. This is a hypothetical illustration, not an LS-Labs observation.
Exposure and selection do not read quite the same information. Microsoft Foundry lets developers set search-only text for each tool; that text is indexed, but models do not see it in MCP responses. Making a tool easier to find therefore does not necessarily change which tool the model finally picks.
How TSO differs from GEO and MCAC
GEO and MCAC look at how AI answers questions. GEO studies and adjusts how visible content is in generative AI answers. MCAC separately checks brand mentions, source citations, factual accuracy and conversion-readiness. TSO looks at whether an AI agent, when it acts, finds and selects your tool, API or MCP server.
The results do not always agree. A brand can appear often in AI answers while agents never call its API, and the reverse can happen too.
Scope and limits of the framework
TSO is an operational framework LS-Labs uses to break the problem down. It is not a ranking factor published by any AI platform. How search ranks tools, how many tools load at once and which model runs are mostly decided by the platform and the agent's developers; what a tool provider can adjust directly is mainly the tool's name, description and parameter schema.
When evaluating a tool-definition change, record candidate retrieval, first choice and execution separately. Compare fixed tasks, candidates and model settings to identify which stage changed.
Nor is TSO about maximizing selection. If a tool cannot do what the user asked, not being chosen is the right outcome. In experiments by Li et al., when only one parameter schema satisfied the user's explicit requirement, all five models chose the compatible tool regardless of their position preferences.
Not the same as Synopsys TSO.ai
Synopsys' TSO.ai stands for Test Space Optimization AI, an AI application for semiconductor testing. It only shares the abbreviation with this entry's TSO.
Sources and scope
- Anthropic: Tool search tool
- OpenAI: Tool search
- Balaji & Sathe, Microsoft (2026-07-29): Tool search
- Model Context Protocol: Specification 2025-06-18, Tools
- Li, Qiao & Gao (2026): Which Metadata Matters for LLM Tool Selection in MCP?
- Blankenstein et al. (2025): BiasBusters
- Synopsys: TSO.ai
Sources as checked for the source-language entry on 2026-10-26; this English edition adds no new claims.