Applied AI
An agent tool description can expose an internal default
Inspect generated callable-object tool descriptions before they reach a model, especially when dataclass defaults include internal settings.
October 28, 2026
An agent's tool description tells the model when and how to use a capability. A reported Agents SDK issue found that wrapping a callable object can pick up an inherited class docstring or a generated dataclass signature instead of the intended __call__ description. In the reproduction, a dataclass description included an internal base URL default. The wrong text is created before a model request is sent.
Inspect the tool contract as model input
During registration, print or assert the final tool name, description, and parameter schema in a safe local test. Compare them with the text approved for model exposure. Include callable classes with abstract bases, dataclasses, named tuples, and plain __call__ methods if your application uses them. The SDK tool guide says docstrings normally provide tool descriptions and that an explicit override is available. What matters is the final generated contract, not the developer's intention.
Remove accidental disclosure
If a generated description contains internal hostnames, defaults, or irrelevant inherited prose, provide a deliberate description override or use a simple function wrapper. Do not put secrets in docstrings or constructor defaults in the first place. Treat tool metadata as text that may be sent to the model provider and may affect tool selection. A nonsensical description can cause a reliability problem even without a secret.
Make registration a checked boundary
Add a small test that fails when the final description changes unexpectedly after a dependency upgrade. Review the issue status and version before assuming the SDK will choose __call__ documentation. This check belongs at agent construction, where the contract is visible and no business action has happened yet.
Adovo’s view
The model acts on the contract it receives, not the one a developer thought they wrote. Tool descriptions are part of the application interface. Inspect them with the same care as an API schema before exposing the agent to real work.