Skip to main content

Agent

adk_agent

Trait Agent 

Source 
pub trait Agent: Send + Sync {
    // Required methods
    fn name(&self) -> &str;
    fn description(&self) -> &str;
    fn sub_agents(&self) -> &[Arc<dyn Agent>];
    fn run<'life0, 'async_trait>(
        &'life0 self,
        ctx: Arc<dyn InvocationContext>,
    ) -> Pin<Box<dyn Future<Output = Result<Pin<Box<dyn Stream<Item = Result<Event, AdkError>> + Send>>, AdkError>> + Send + 'async_trait>>
       where 'life0: 'async_trait,
             Self: 'async_trait;

    // Provided methods
    fn interaction_mode(&self) -> AgentInteractionMode { ... }
    fn supports_agent_transfer(&self) -> bool { ... }
    fn capabilities(&self) -> AgentCapabilities { ... }
    fn topology(&self) -> Option<AgentTopology> { ... }
    fn configure_run(&self, _agent_name: &str, _config: &mut RunConfig) { ... }
    fn transfer_targets_for(&self, _agent_name: &str) -> Option<Vec<String>> { ... }
    fn strict_transfer_policy(&self) -> bool { ... }
    fn govern_transfer<'life0, 'life1, 'async_trait>(
        &'life0 self,
        _request: &'life1 AgentTransferRequest,
    ) -> Pin<Box<dyn Future<Output = Result<AgentTransferDecision, AdkError>> + Send + 'async_trait>>
       where 'life0: 'async_trait,
             'life1: 'async_trait,
             Self: 'async_trait { ... }
}
Expand description

The fundamental trait for all ADK agents.

Every agent — whether a simple LLM wrapper, a multi-step workflow, or a composite orchestrator — implements this trait. The runtime invokes run with an InvocationContext and consumes the returned EventStream.

Required Methods§

Source

fn name(&self) -> &str

Returns the unique name of this agent.

Source

fn description(&self) -> &str

Returns a human-readable description of this agent’s purpose.

Source

fn sub_agents(&self) -> &[Arc<dyn Agent>]

Returns the child agents managed by this agent.

Source

fn run<'life0, 'async_trait>( &'life0 self, ctx: Arc<dyn InvocationContext>, ) -> Pin<Box<dyn Future<Output = Result<Pin<Box<dyn Stream<Item = Result<Event, AdkError>> + Send>>, AdkError>> + Send + 'async_trait>>
where 'life0: 'async_trait, Self: 'async_trait,

Executes the agent and returns a stream of events.

Provided Methods§

Source

fn interaction_mode(&self) -> AgentInteractionMode

Returns the agent’s primary interaction pattern.

Existing agents remain request/response by default. Realtime agents override this without introducing a new atomic agent kind or changing composition semantics.

Source

fn supports_agent_transfer(&self) -> bool

Whether this agent participates in LLM-driven agent transfer and may be resumed directly across conversation turns.

When a session persists across turns, the runner inspects history to decide which agent should handle the next user message. LLM-based and custom agents return the default true, so the runner can hand a new turn back to whichever agent responded last.

Deterministic workflow agents (sequential, parallel, loop, conditional) override this to return false. Their sub-agents must not be resumed individually: doing so would skip the workflow’s other sub-agents on subsequent turns. Returning false makes the runner resume the workflow root instead, so every sub-agent runs again on each turn.

Source

fn capabilities(&self) -> AgentCapabilities

Declares the execution-plane capabilities this agent supports.

The default remains compatible with existing custom agents: transfer, shared-state, cancellation, and request metadata follow the established Agent and InvocationContext contracts. Runtime tool injection and exact-call relationship confirmation are opt-in because an implementation must actively consume those facilities.

Source

fn topology(&self) -> Option<AgentTopology>

Returns portable composition metadata when this agent owns an explicit topology.

Leaf agents and legacy composites return None. The default keeps this additive API backward compatible for existing Agent implementations.

Source

fn configure_run(&self, _agent_name: &str, _config: &mut RunConfig)

Applies agent-composition policy to a run before the runtime creates the invocation context for agent_name.

Composite agents can use this hook to inject invocation-scoped tools, constrain transfer targets, or tighten depth and concurrency limits. The default is a no-op, preserving the behavior of existing agents.

Source

fn transfer_targets_for(&self, _agent_name: &str) -> Option<Vec<String>>

Returns the exact handoff targets allowed for agent_name, when this agent owns an explicit transfer topology.

None asks the runtime to retain its legacy parent/peer discovery. Some(vec![]) explicitly forbids handoff from the named agent.

Source

fn strict_transfer_policy(&self) -> bool

Whether transfer-policy violations should fail the run.

Legacy agent trees return false, so missing targets and exceeded depth retain their historical warn-and-stop behavior. Validated composites return true to make topology violations observable errors.

Source

fn govern_transfer<'life0, 'life1, 'async_trait>( &'life0 self, _request: &'life1 AgentTransferRequest, ) -> Pin<Box<dyn Future<Output = Result<AgentTransferDecision, AdkError>> + Send + 'async_trait>>
where 'life0: 'async_trait, 'life1: 'async_trait, Self: 'async_trait,

Applies asynchronous policy to an otherwise valid transfer.

Runner calls this after exact target validation and before control moves to the target. Ordinary agents allow transfers, preserving legacy behavior; validated composites can attach authorization, lifecycle hooks, audit logging, or other policy without teaching Runner their schema.

Dyn Compatibility§

This trait is dyn compatible.

In older versions of Rust, dyn compatibility was called "object safety".

Implementors§