# AI Agent Distribution The practice of packaging, sharing, and deploying complete [[AI Agents]] beyond their original context. An agent is more than a skill: it bundles identity, memory configuration, skill sets, routing rules, and behavioral guidelines into a coherent role. ## What gets distributed An agent package includes: - **Identity** ([[AI Agent Identity]]): role definition, personality, expertise boundaries, behavioral rules - **Skill manifest**: which [[AI Agent Skills]] the agent uses and how they're combined - **Memory configuration**: what the agent remembers and how ([[AI Agent Memory]]) - **Routing rules**: when this agent should be activated ([[AI Agent Routing]]) - **Harness configuration**: how the agent integrates with the [[AI Agent Harness]] ## Distribution scopes - **Personal**: reusing agents across projects (e.g., a code reviewer agent that works on any codebase) - **Team**: sharing specialized agents within a team (e.g., a team's PR review agent with team-specific conventions) - **Enterprise**: deploying governed agent configurations across the organization with approved skill sets and compliance guardrails - **Community**: publishing agent templates and configurations for others to adapt ## Current state Agent distribution today is mostly manual: - Agent definitions live as Markdown files (SOUL.md, DEPENDENCIES.md) in vault or repo structures - [[Claude Code Plugins]] support agent definitions with `.claude/agents/` conventions - Sharing means copying agent folders or publishing repos - No standard format for agent packaging across tools ## Key challenges - **Context dependency**: agents are deeply tied to their context (vault structure, available tools, domain knowledge). A "ghostwriter agent" trained on your writing style doesn't port to someone else - **Skill dependency management**: agents compose multiple skills. Distributing an agent means distributing its skill dependencies too - **Identity portability**: agent identity definitions are tool-specific. An agent defined for Claude Code doesn't work in Cursor without adaptation - **Trust**: agents with tool access and autonomy are high-risk to run from untrusted sources. The [[Lethal Trifecta for AI Agents]] applies doubly to distributed agents ## The portable agent future The ecosystem is moving toward: - Standardized agent definition formats - Agent registries with trust levels and capability declarations - Protocols like [[Agent Client Protocol (ACP)]] for cross-platform agent communication - Separation of agent identity (portable) from agent runtime (tool-specific) ## References - ## Related - [[AI Agents]] - [[AI Skill Distribution]] - [[AI Skill Composability]] - [[AI Agent Identity]] - [[AI Agent Memory]] - [[AI Agent Skills]] - [[AI Agent Routing]] - [[AI Agent Harness]] - [[Claude Code Plugins]] - [[Agent Client Protocol (ACP)]] - [[Lethal Trifecta for AI Agents]] - [[Agentic Engineering]]