Create an Agent Definition — agent

AgentDefinition describes a specialized agent that can be used by a lead agent to delegate tasks. It bundles together a system prompt, tools, and metadata about what the agent can do.

Usage

agent_definition(
  name,
  description,
  prompt,
  tools = list(),
  model = "inherit",
  skills = list(),
  disallowed_tools = NULL,
  memory = NULL,
  mcp_servers = NULL,
  initial_prompt = NULL,
  max_requests = NULL,
  permission_mode = NULL
)

Arguments

name: Unique name for this agent type
description: Brief description of what this agent does (shown to lead agent)
prompt: System prompt for this agent
tools: Optional list of tools for this agent
model: Model to use (default: "inherit" uses parent's model)
skills: Optional list of skills to load
disallowed_tools: Optional tool denylist for this sub-agent
memory: Optional memory text appended to this sub-agent's prompt
mcp_servers: Optional MCP server names to load for this sub-agent
initial_prompt: Optional text prepended to delegated tasks
max_requests: Optional non-negative whole-number sub-agent request limit
permission_mode: Optional permission mode. A child may keep the lead mode or narrow it to "readonly"; a "full" lead may select any mode. Use disallowed_tools and max_requests for additional limits.

Value

An AgentDefinition object

Examples

if (FALSE) { # \dontrun{
# Define a code review agent
code_reviewer <- agent_definition(
  name = "code_reviewer",
  description = "Reviews code for bugs, style issues, and best practices",
  prompt = "You are an expert code reviewer...",
  tools = list(tool_read_file, tool_list_files)
)

# Use with a lead agent
lead <- LeadAgent$new(
  chat = ellmer::chat("openai/gpt-4o"),
  sub_agents = list(code_reviewer)
)
} # }