The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references external URLs for the purpose of downloading and setting up software.
Fetches the core framework from a GitHub repository under the Nous Research organization.
References a third-party 'Orange Book' guide from a personal GitHub repository (alchaincyf/hermes-agent-orange-book), which is an unverifiable community resource.
[COMMAND_EXECUTION]: Provides specific commands for users to set up and run the agent framework locally.
Includes commands for git clone https://github.com/NousResearch/hermes-agent.git and installing packages with pip install -r requirements.txt or poetry install.
Demonstrates interactive usage via the CLI with commands like python -m hermes_agent and hermes chat.
[INDIRECT_PROMPT_INJECTION]: The described framework has a high-privilege capability set that can be exploited by indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through standard chat interfaces (agent.chat()) and multi-agent workflows.
Boundary markers: The provided documentation lacks examples of sanitization or delimiters to prevent the agent from obeying instructions embedded in retrieved data.
Capability inventory: The framework includes tools such as CodeExecutorTool (subprocess execution), FileSystemTool (file read/write), and GitTool (repository management).
Sanitization: Although tools have theoretical configuration options (e.g., allowed_paths), the primary task of interpreting natural language instructions from users or external sources (via WebSearchTool) presents a persistent injection risk.
[METADATA_POISONING]: The skill uses misleading or speculative information in its descriptive text.
The guide claims the framework was released in 'February 2026'. As this is a future date, the content is potentially hallucinatory or deceptive regarding the framework's actual availability, versioning, and features.

hermes-agent-framework