MedGaugeCapabilities / B · Embodiment (tools)
B.3 Tool invocation Active
Given a task and a set of available tools, the agent calls the right tool at the right point with valid arguments. Graded on the trajectory: whether each required tool was called, whether arguments were valid, and the count of unnecessary calls.
Evidence: none. No benchmark card lists this capability yet. No level is issued: methodology 0.1.0 sets no level thresholds.
- Domain
- B · Embodiment (tools)
- Evidence form
- program
- Veto constraints
- —
- Depends on
- —
- Required by
- B.4B.6C.5D.1D.10D.2D.6
- Internal task families
- —
- External anchors
- tau-bench
- Version
- 0.1.0 · 0.1.0
Definition in detail
Reliable agents look facts up instead of recalling them and call each tool with arguments drawn from the right record. Unneeded calls add cost and, for write-capable tools, risk.
Inputs a task gives
- A task statement with its static materials
- A catalogue of available tools with their schemas
- An initial environment state, such as a sandboxed record store
Artifact required
The end-state artifact the task requires, plus the recorded trajectory of tool calls with their arguments.
How it is graded
A program checks the trajectory: whether each required tool was called before the step that depends on it, whether arguments validate against the tool schema and match the reference constraints, and how many unnecessary calls were made. The end state is checked separately by the task's own grader.
In scope
- Selecting the right tool for each step
- Supplying valid arguments drawn from the correct record
- Avoiding redundant or unneeded calls
Out of scope
- Handling error, empty and timed-out returns - graded under B.4
- Assembling the tool set into a harness - graded under B.6
- Calls requested by instructions planted in the data - X.4
Typical failures
Expected failure patterns; not measured rates.
- Answering from memory where the task expects a lookup tool to be called
- Taking arguments from the wrong record or encounter
- Repeating identical calls
- Using a write-capable tool where a read-only one was sufficient
Standards: std-mcp-2026-07-28 · std-atif
Benchmarks
No benchmark card lists this capability.
Grading references
No grading reference is mapped to this capability.