The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The orchestration framework processes natural language goals and tool outputs to generate and execute task graphs automatically.
Ingestion points: User goals passed to execution methods and data retrieved through tools such as file_read, grep, or external Model Context Protocol (MCP) servers.
Boundary markers: The documentation does not demonstrate the use of delimiters or specific instructions to prevent agents from following commands embedded in processed files or tool outputs.
Capability inventory: The framework grants agents access to bash for shell execution and file_write/file_edit for modifying the local file system.
Sanitization: Zod is used for validating final structured results, but no mechanism for sanitizing agent-generated tool arguments is described.
[COMMAND_EXECUTION]: The framework features a built-in bash tool, enabling agents to execute arbitrary shell commands to perform development and system tasks.
[EXTERNAL_DOWNLOADS]: The instructions include downloading the framework's core package from a registry and fetching MCP servers using the npx utility.
[DYNAMIC_EXECUTION]: The framework supports Model Context Protocol integration, which involves spawning external processes at runtime and executing tools provided by those servers.

open-multi-agent-orchestration