Struct AgentConfig 

pub struct AgentConfig {
    pub max_turns: Option<usize>,
    pub max_tokens: Option<u32>,
    pub system_prompt: String,
    pub model: String,
    pub retry: RetryConfig,
    pub streaming: bool,
    pub tool_timeout_ms: Option<u64>,
    pub usage_limits: Option<UsageLimits>,
    pub max_parallel_tools: Option<usize>,
}

Configuration for the agent loop

Fields

max_turns: Option<usize>

Maximum number of turns (LLM round-trips) before stopping

max_tokens: Option<u32>

Maximum tokens per response.

If None, the SDK uses the provider/model-specific default.

system_prompt: String

System prompt for the agent

model: String

Model identifier

retry: RetryConfig

Retry configuration for transient errors

streaming: bool

Enable streaming responses from the LLM.

When true, emits TextDelta and ThinkingDelta events as text arrives in real-time. When false (default), waits for the complete response before emitting Text and Thinking events.

tool_timeout_ms: Option<u64>

Optional per-tool execution timeout in milliseconds.

When set, the agent loop races each tool’s execute() future against this budget at the SDK boundary (mirroring SubagentConfig::timeout_ms). A tool that exceeds the budget is stopped and reported with a synthetic timeout ToolResult, keeping the tool_use / tool_result history balanced even for non-cooperative tools. None (default) disables the boundary timeout entirely.

usage_limits: Option<UsageLimits>

Optional run-level token / cost budgets.

When set, the agent loop checks the cumulative token usage (and the estimated USD cost, when the provider/model has pricing metadata) at every turn-continuation boundary. If a configured limit is exceeded the run stops with AgentRunState::BudgetExceeded / TurnOutcome::BudgetExceeded instead of starting another turn. None (default) disables budgeting entirely.

max_parallel_tools: Option<usize>

Maximum number of read-only (ToolTier::Observe) tool calls the SDK runs concurrently within a single parallel batch.

None (default) keeps the historical unbounded behavior — every adjacent observe-tier call in a turn is dispatched at once. Some(1) forces strictly sequential execution; Some(n) caps the in-flight count at n. Some(0) is not meaningful and is treated as Some(1) (sequential). Result ordering is always preserved.

Trait Implementations

impl Clone for AgentConfig

fn clone(&self) -> AgentConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for AgentConfig

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

Formats the value using the given formatter.

impl Default for AgentConfig

fn default() -> AgentConfig

Returns the “default value” for a type.