The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The Agentica infrastructure API specification defines a high-surface-area environment for multi-agent coordination, which is susceptible to indirect prompt injection where data from one agent influences another.
Ingestion Points: The framework includes BlackboardCache (blackboard.py) for message passing between agents, SharedContext (claude_scope.py) for file operation logging and caching, and CoordinationDB (coordination.py) for tracking broadcasts and tasks. These components ingest data from multiple agents that may be processing untrusted external content.
Boundary Markers: The documentation does not specify the use of delimiters or 'ignore embedded instructions' warnings for data retrieved from the shared blackboard or memory services.
Capability Inventory: The create_claude_scope function in claude_scope.py explicitly includes high-privilege capabilities such as bash, write_file, and edit_file in the agent execution environment, which could be targeted by injected instructions.
Sanitization: The specification lacks explicit details on sanitization, validation, or escaping of the natural language (prose) or formal content passed through the SharedContextEntry or HandoffState structures.
[COMMAND_EXECUTION]: The documentation for claude_scope.py indicates the provision of a bash tool within the standard execution scope for agents using this infrastructure, representing a significant capability surface for command execution.

agentica-infrastructure