Struct MultiAgentConfig 

pub struct MultiAgentConfig {
    pub enabled: bool,
    pub max_sub_agents: usize,
    pub max_agent_depth: i32,
    pub child_permission_mode: ChildPermissionMode,
    pub child_excluded_tools: Vec<String>,
    pub child_reasoning_effort: Option<ReasoningEffort>,
    pub child_read_only: bool,
}

Configuration for the multi-agent subsystem.

Default

Multi-agent is enabled by default with limits of 8 sub-agents and depth 1. Disable it with:

use agent_works::AgentBuilder;

let agent = AgentBuilder::new(client)
    .without_multi_agent()
    .build()?;

Limits

Limits only take effect when enabled is true.

Fields

enabled: bool

Enable multi-agent capabilities.

When true (default): all 6 multi-agent tools are registered, and the main agent’s system prompt is augmented with usage guidance.

When false: no multi-agent tools are registered, and the system prompt does not mention multi-agent capabilities.

max_sub_agents: usize

Maximum number of concurrent sub-agents.

Default: 8. Only enforced when enabled is true.

max_agent_depth: i32

Maximum nesting depth for sub-agents.

Default: 1 (only direct children of root can spawn). A value of 0 means no sub-agents can be spawned at all. Only enforced when enabled is true.

child_permission_mode: ChildPermissionMode

Permission mode for spawned child agents.

Default: ChildPermissionMode::Full.

child_excluded_tools: Vec<String>

Business tool names to EXCLUDE from child agents.

By default children inherit every business tool registered on the parent. Some tools are inherently root-level (e.g. decompose/merge orchestration): a leaf agent that lacks spawn_agent should not be handed a “split this into parallel sub-agents” tool — it would plan work it cannot execute. Mark such tools here so build_child_runtime skips them, and children do the work inline instead of trying to orchestrate further.

child_reasoning_effort: Option<ReasoningEffort>

Optional reasoning-effort override applied to every child agent.

Child agents do narrow, focused slices, so they rarely need the parent’s full reasoning depth — and on reasoning-heavy models (deepseek-v4-pro) an unbounded child can “think” itself into a runaway (30KB+ of reasoning with no tool call or answer). Setting e.g. ReasoningEffort::Low caps that cost. None (default) leaves the child on the framework default.

child_read_only: bool

Whether child agents are recommended to be read-only.

The framework is domain-agnostic: it cannot classify which business tools mutate state and which merely read, so this is a prompt-level suggestion only — it does not remove or gate any tool. When true (default), [MultiAgentRuntime::build_child_runtime] appends a read-only nudge to every child’s system prompt, telling it to investigate and report rather than mutate.

For a hard guarantee, exclude mutating tools at the business layer via MultiAgentConfig::child_excluded_tools; the framework only suggests read-only, it cannot enforce it. Set this to false when children are meant to write (the codex-style symmetric model).

Implementations

impl MultiAgentConfig

pub fn enabled() -> MultiAgentConfig

Create a config with multi-agent enabled and default limits.

pub fn with_limits( max_sub_agents: usize, max_agent_depth: i32, ) -> MultiAgentConfig

Create a config with multi-agent enabled and custom limits.

Trait Implementations

impl Clone for MultiAgentConfig

fn clone(&self) -> MultiAgentConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for MultiAgentConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.

impl Default for MultiAgentConfig

fn default() -> MultiAgentConfig

Returns the “default value” for a type.