MedGaugeCapabilities / B · Embodiment (tools)
B.1 Tool creation Active
Given a requirement, the agent writes a runnable tool (script, function or MCP server) with its own tests. Graded by passing a supplied test set and by conformance of the tool interface to the given schema.
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
- C.6 Deterministic program implementation
- Required by
- B.2B.5
- Internal task families
- —
- External anchors
- SkillsBench
- Version
- 0.1.0 · 0.1.0
Definition in detail
An agent that can write a small, tested tool for a repeated operation makes later runs deterministic and auditable instead of re-deriving the operation each time.
Inputs a task gives
- A requirement statement for the tool
- An interface schema for inputs and outputs
- Sample inputs with expected outputs
- Runtime constraints such as language and permitted dependencies
Artifact required
Tool source (script, function or MCP server), the agent's own test file, and a manifest with name, entry_point, input_schema, output_schema and dependencies.
How it is graded
A program runs the supplied test set against the tool in a clean environment and validates the interface against the given schema. The agent's own tests must run and pass. A tool that reads expected outputs from the test set or fails on undeclared dependencies fails.
In scope
- Implementing the stated requirement behind the given interface
- Writing tests that exercise the requirement, including edge cases
- Declaring dependencies and entry points accurately
Out of scope
- Writing the registry entry for the tool - graded under B.2
- Implementing a written scoring or calculation rule against a hidden test set - graded under C.6
- Revising an existing tool from trajectory evidence - graded under B.5
Typical failures
Expected failure patterns; not measured rates.
- Tests that cover only the sample cases or restate the implementation
- Hard-coding sample outputs instead of implementing the rule
- Field names or types that drift from the interface schema
- Relying on packages present in the authoring environment but not declared
Standards: std-mcp-2026-07-28
Benchmarks
No benchmark card lists this capability.
Grading references
No grading reference is mapped to this capability.