reson-agentic

Agents are just functions - production-grade LLM agent framework for Rust.

Features

• Multi-provider support: Anthropic, OpenAI, Google Gemini, OpenRouter, AWS Bedrock
• Native tool calling with structured outputs via #[derive(Tool)]
• Agent macro for ergonomic agent definitions with #[agentic]
• Streaming responses with reasoning/thinking support
• Google File API for video/large media uploads
• Retry with exponential backoff
• Clone-friendly clients for use in async contexts

Installation

[dependencies]
reson-agentic = "0.1"
tokio = { version = "1", features = ["full"] }
serde = { version = "1", features = ["derive"] }

Quick Start

Basic Chat

use reson_agentic::providers::{GoogleGenAIClient, GenerationConfig, InferenceClient};
use reson_agentic::types::ChatMessage;
use reson_agentic::utils::ConversationMessage;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let client = GoogleGenAIClient::new("your-api-key", "gemini-2.0-flash");

    let messages = vec![
        ConversationMessage::Chat(ChatMessage::user("Hello!"))
    ];

    let config = GenerationConfig::new("gemini-2.0-flash");
    let response = client.get_generation(&messages, &config).await?;

    println!("{}", response.content);
    Ok(())
}

Tool Definitions with #[derive(Tool)]

Define type-safe tools that automatically generate JSON schemas for LLM function calling:

use reson_agentic::Tool;
use serde::{Deserialize, Serialize};

/// Search the web for information
#[derive(Tool, Serialize, Deserialize, Debug)]
struct WebSearch {
    /// The search query to execute
    query: String,
    /// Maximum number of results to return
    max_results: Option<u32>,
}

/// Get weather for a location
#[derive(Tool, Serialize, Deserialize, Debug)]
struct GetWeather {
    /// City name or coordinates
    location: String,
    /// Temperature unit: "celsius" or "fahrenheit"
    unit: Option<String>,
}

// Access generated schema
let schema = WebSearch::schema();       // JSON Schema object
let name = WebSearch::tool_name();      // "web_search"
let desc = WebSearch::description();    // "Search the web for information"

The #[derive(Tool)] macro:

• Converts struct name to snake_case for the tool name
• Uses doc comments as descriptions (struct doc -> tool description, field docs -> parameter descriptions)
• Generates proper JSON Schema with types, required fields, and array items
• Supports String, bool, i32/i64/u32/u64, f32/f64, Vec<T>, and Option<T>

Agent Functions with #[agentic]

The #[agentic] macro transforms an async function into an agent. It:

1. Creates a Runtime automatically and injects it into the function
2. Validates that runtime.run() or runtime.run_stream() is called
3. Configures the model from the macro attribute

use reson_agentic::agentic;
use reson_agentic::runtime::{Runtime, ToolFunction};
use reson_agentic::error::Result;

/// Analyze text and answer questions
#[agentic(model = "gemini:gemini-2.0-flash")]
async fn analyze_text(
    text: String,
    question: String,
    runtime: Runtime,  // Injected by macro - callers don't pass this
) -> Result<serde_json::Value> {
    // Register tools with the runtime
    runtime.register_tool_with_schema(
        WebSearch::tool_name(),
        WebSearch::description(),
        WebSearch::schema(),
        ToolFunction::Sync(Box::new(|args| {
            let query = args["query"].as_str().unwrap_or("");
            Ok(format!("Search results for: {}", query))
        })),
    ).await?;

    // Run the agent - runtime is mutable internally
    runtime.run(
        Some(&format!("Text: {}\n\nQuestion: {}", text, question)),
        Some("You are a helpful assistant. Use tools when needed."),
        None,  // history
        None,  // output_type
        None,  // temperature
        None,  // top_p
        None,  // max_tokens
        None,  // model override
        None,  // api_key override
    ).await
}

// Call the agent - runtime parameter is NOT passed by caller
let result = analyze_text(
    "The quick brown fox...".to_string(),
    "What animal is mentioned?".to_string(),
).await?;

Video/Media Upload (Google Gemini)

Upload and analyze videos using Google's File API:

use reson_agentic::providers::{GoogleGenAIClient, FileState};
use reson_agentic::types::{ChatRole, MediaPart, MediaSource, MultimodalMessage};

let client = GoogleGenAIClient::new(api_key, "gemini-2.0-flash");

// Upload video
let video_bytes = std::fs::read("video.mp4")?;
let uploaded = client.upload_file(&video_bytes, "video/mp4", Some("my-video")).await?;

// Wait for processing (required for videos)
if uploaded.state == FileState::Processing {
    client.wait_for_file_processing(&uploaded.name, Some(120)).await?;
}

// Create multimodal message
let message = MultimodalMessage {
    role: ChatRole::User,
    parts: vec![
        MediaPart::Video {
            source: MediaSource::FileUri {
                uri: uploaded.uri.clone(),
                mime_type: Some("video/mp4".to_string()),
            },
            metadata: None,
        },
        MediaPart::Text { text: "Describe this video".to_string() },
    ],
    cache_marker: None,
};

// Clean up when done
client.delete_file(&uploaded.name).await?;

Supported Media Types

Providers

*Requires google-adc feature: reson-agentic = { version = "0.1", features = ["google-adc"] }

All clients implement Clone for easy use in async contexts.

Examples

See the examples directory:

• video_upload.rs - Video analysis with Google Gemini and #[agentic] macro
• simple_tools.rs - Basic tool registration and execution
• tool_call_chain.rs - Multi-turn tool calling
• dynamic_tool_parsing.rs - Type-safe tool parsing with Deserializable
• templating_example.rs - Prompt templates with minijinja
• store_usage.rs - Context storage patterns

Run examples with:

GOOGLE_GEMINI_API_KEY=your_key cargo run --example video_upload -- video.mp4

Feature Flags

[dependencies]
reson-agentic = { version = "0.1", features = ["full"] }

License

Apache-2.0