Crate llm_agent_runtime 

agent-runtime

A unified Tokio agent runtime that brings together orchestration, memory, knowledge graph, and a ReAct agent loop in a single crate.

Feature Flags

Feature	Content
orchestrator (default)	Circuit breaker, retry, dedup, backpressure, pipeline
memory (default)	Episodic, semantic, working memory + decay
graph (default)	In-memory knowledge graph (BFS/DFS/shortest-path)
wasm	ReAct agent loop, tool registry
full	All of the above

Quick Start

use llm_agent_runtime::prelude::*;

// build() is now infallible — the typestate pattern enforces that
// with_agent_config() is called before build() at compile time.
let runtime = AgentRuntime::builder()
    .with_agent_config(AgentConfig::new(5, "my-model"))
    .build();

// run_agent is async; drive it with your async runtime.
// tokio::runtime::Runtime::new().unwrap().block_on(async {
//     let session = runtime
//         .run_agent(AgentId::new("agent-1"), "Hello", |_ctx: String| async {
//             "Thought: done\nAction: FINAL_ANSWER hi".to_string()
//         })
//         .await
//         .unwrap();
//     println!("Steps: {}", session.step_count());
// });

Extension Points

The primary extension point is the [PersistenceBackend] trait (enabled by the persistence feature). Implement it to persist agent sessions and per-step checkpoints to any storage backend — disk, Redis, S3, or a database. The bundled [FilePersistenceBackend] stores each checkpoint as a <key>.bin file in a local directory and serves as a reference implementation.

Additional extension points:

• [LlmProvider] (providers feature) — plug in any LLM API.
• CircuitBreakerBackend (redis-circuit-breaker feature) — distributed circuit-breaker state via Redis.

Upstream Crates

This crate implements standalone versions of the APIs from:

• tokio-prompt-orchestrator (Mattbusel/tokio-prompt-orchestrator)
• tokio-agent-memory (Mattbusel/tokio-agent-memory)
• mem-graph (Mattbusel/mem-graph)
• wasm-agent (Mattbusel/wasm-agent)
• tokio-memory (Mattbusel/tokio-memory)

Re-exports

pub use error::AgentRuntimeError;

pub use runtime::AgentRuntime;

pub use runtime::AgentRuntimeBuilder;

pub use runtime::AgentSession;

pub use types::AgentId;

pub use types::MemoryId;

pub use memory::MemoryItem;

pub use graph::Entity;

pub use graph::EntityId;

pub use graph::GraphStore;

pub use graph::Relationship;

pub use orchestrator::CircuitBreaker;

pub use orchestrator::Pipeline;

pub use orchestrator::RetryKind;

pub use orchestrator::RetryPolicy;

pub use agent::Action;

pub use agent::AgentConfig;

pub use agent::ReActLoop;

pub use agent::ReActStep;

pub use agent::ToolSpec;

pub use agent::ToolValidator;

pub use metrics::MetricsSnapshot;

Modules

agent

Module: Agent

error

Unified error type for the agent-runtime crate.

graph

Module: Graph

memory

Module: Memory

metrics

Module: Metrics

orchestrator

Module: Orchestrator

prelude

Convenience re-exports for the most commonly used types.

runtime

Module: AgentRuntime

types

Core identifier types, available without feature gates.

util

Shared utility functions used across modules.