Struct AgentBuilder 

pub struct AgentBuilder { /* private fields */ }

Builder for constructing an Agent with custom configuration.

Supports two modes:

1. Standalone – the builder loads the model itself (simple, one agent):

let agent = AgentBuilder::new()
    .backend(Backend::Cpu)
    .model_path("model.gguf")
    .build()
    .expect("Failed to build agent");

2. Shared engine – multiple agents share one model (no redundant loading):

let engine = Arc::new(InferenceEngine::load(InferenceConfig {
    backend: Backend::Vulkan,
    model_path: "model.gguf".into(),
    n_gpu_layers: 99,
    ..Default::default()
}).unwrap());

let agent_a = AgentBuilder::new()
    .engine(engine.clone())
    .system_prompt("You are agent A.")
    .build().unwrap();

let agent_b = AgentBuilder::new()
    .engine(engine.clone())
    .system_prompt("You are agent B.")
    .build().unwrap();

Implementations

impl AgentBuilder

pub fn new() -> Self

pub fn engine(self, engine: Arc<InferenceEngine>) -> Self

Use a shared InferenceEngine instead of loading a new model.

When set, backend(), model_path(), n_gpu_layers(), dll_*(), cache_dir(), and app_name() are ignored — the engine already has those configured.

pub fn backend(self, backend: Backend) -> Self

Set the compute backend (CPU, CUDA, Vulkan, etc.)

pub fn model_path(self, path: &str) -> Self

Path to the GGUF model file.

pub fn n_gpu_layers(self, n: i32) -> Self

Number of layers to offload to GPU (-1 = all, 0 = none).

pub fn app_name(self, name: &str) -> Self

Application name (used for cache directory).

pub fn cache_dir(self, dir: PathBuf) -> Self

Directory for caching downloaded DLLs.

pub fn explicit_dll_path(self, path: PathBuf) -> Self

Explicit path to the llama.cpp DLL (bypasses download).

pub fn dll_version(self, version: &str) -> Self

DLL version tag to download.

pub fn chat_template(self, template: &str) -> Self

Set a custom chat template (Jinja).

pub fn system_prompt(self, prompt: &str) -> Self

System prompt that instructs the model on its role and tool usage.

pub fn n_ctx(self, n: u32) -> Self

Context window size in tokens.

pub fn max_iterations(self, n: usize) -> Self

Maximum agent loop iterations (0 = unlimited).

pub fn max_tokens_per_completion(self, n: usize) -> Self

Maximum tokens per model completion.

pub fn temperature(self, temp: f32) -> Self

Sampling temperature.

pub fn top_k(self, k: i32) -> Self

Top-K sampling parameter.

pub fn min_p(self, p: f32) -> Self

Min-P sampling parameter.

pub fn repeat_penalty(self, p: f32) -> Self

Repetition penalty.

pub fn stop_sequence(self, stop: &str) -> Self

Add a stop sequence.

pub fn auto_approve(self) -> Self

Auto-approve all tool calls (YOLO mode — dangerous!).

pub fn permission_callback( self, cb: impl Fn(&PermissionRequest) -> PermissionDecision + Send + Sync + 'static, ) -> Self

Set a permission callback for interactive approval.

pub fn tool(self, tool: Box<dyn Tool>) -> Self

Add a custom tool.

pub fn skip_builtin_tools(self) -> Self

Skip registering built-in tools (bash, read, write, edit, glob).

pub fn no_skills(self) -> Self

Disable skill discovery entirely.

pub fn skills_path(self, path: PathBuf) -> Self

Add an extra directory to search for skills.

pub fn activate_skill(self, name: &str) -> Self

Explicitly activate a skill by name.

pub fn no_agents_md(self) -> Self

Disable AGENTS.md discovery entirely.

pub fn scheduler(self, scheduler: Arc<InferenceScheduler>) -> Self

Set an inference scheduler to limit concurrent inferences.

Use InferenceScheduler::new(1) to serialize all inference (one agent at a time), or a higher value for controlled parallelism. Without a scheduler, agents run truly parallel (safe, but GPU-heavy).

pub fn build(self) -> Result<Agent, AgentError>

Build the agent.

If a shared engine was provided via .engine(), it is reused. Otherwise, a new engine is created from the standalone fields.

Trait Implementations

impl Default for AgentBuilder

fn default() -> Self

Returns the “default value” for a type.