The Agent Skills Directory

[SAFE]: No malicious patterns or security risks were detected. The skill is a legitimate development framework for AI agent optimization using official Anthropic APIs and well-known libraries.
[EXTERNAL_DOWNLOADS]: The Python reference package includes a requirement for a preview version of the Anthropic SDK hosted on Stainless (app.stainless.com). Stainless is a well-known service for SDK generation used by major technology companies, making this a legitimate source for development preview builds.
[COMMAND_EXECUTION]: The skill implements CLI tools and scripts that manage the creation, execution, and cleanup of Claude Managed Agents. These commands utilize the Anthropic SDK to interact with the platform as intended by the framework's design.
[PROMPT_INJECTION]: The skill provides a surface for indirect prompt injection as it is designed to evaluate agents against external task data.
Ingestion points: Untrusted data enters the agent context through the input_text field of DummyTask objects and associated workspace files, as defined in src/optimizespec/tasks.py.
Boundary markers: The framework uses structured templates (e.g., DEFAULT_SEED_CANDIDATE in src/optimizespec/candidate.py) to define the agent's operating instructions and success criteria.
Capability inventory: The system has capabilities to manage files, perform network operations, and invoke specialist tools via the Managed Agents SDK implementation in src/optimizespec/runtime.py.
Sanitization: Task data is handled as separate file and message entities within the Managed Agents environment to maintain context boundaries.

optimizespec-common