Struct AgentLoopConfig 

pub struct AgentLoopConfig {
    pub max_iterations: usize,
    pub max_tokens_per_completion: usize,
    pub temperature: f32,
    pub top_k: i32,
    pub min_p: f32,
    pub repeat_penalty: f32,
    pub auto_compact: bool,
    pub compaction_threshold_pct: f32,
    pub compaction_keep_recent: usize,
    pub n_batch: usize,
    pub stop_sequences: Vec<String>,
}

Configuration for the agent loop.

Fields

max_iterations: usize

Maximum number of iterations (tool-use rounds). 0 = unlimited.

max_tokens_per_completion: usize

Maximum tokens to generate per completion.

temperature: f32

Temperature for sampling.

top_k: i32

Top-K sampling parameter.

min_p: f32

Min-P sampling parameter.

repeat_penalty: f32

Repetition penalty.

auto_compact: bool

Enable automatic context compaction (default: true).

compaction_threshold_pct: f32

Percentage of context window that triggers compaction (0.0–1.0).

compaction_keep_recent: usize

Number of recent messages to keep when compacting.

n_batch: usize

Maximum tokens per batch during prompt encoding (default: 512).

Large prompts are split into chunks of this size for decoding. Smaller values use less peak GPU memory; larger values may be faster on GPUs with plenty of VRAM.

stop_sequences: Vec<String>

List of stop sequences. If the model generates one of these, it stops.

Trait Implementations

impl Default for AgentLoopConfig

fn default() -> Self

Returns the “default value” for a type.