agentwerk 0.1.3

A minimal Rust crate that gives any application agentic capabilities.
Documentation

Installation

cargo add agentwerk

Quick Start

use agentwerk::{Agent, GlobTool, provider_from_env};

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let (provider, model) = provider_from_env()?;

    let output = Agent::new()
        .provider(provider)
        .model(model)
        .instruction_prompt("Find all Rust source files.")
        .tool(GlobTool)
        .run()
        .await?;

    println!("{}", output.response_raw);
    Ok(())
}

Use Cases

Example applications built with this project.

Consider configuring your LLM provider (see Environment).

Project Scanner

Two phases: a discovery agent finds files worth reading, then a pool of agents summarizes each file in parallel.

make use_case name=project-scanner -- ./

Output:

{
  "languages": ["config", "docs", "rust"],
  "files": [
    {
      "file": "README.md",
      "summary": "Project documentation for agentwerk, a Rust crate for building agentic LLM applications.",
      "language": "docs"
    },
    {
      "file": "crates/agentwerk/src/agent/loop.rs",
      "summary": "Implements the main agent loop that calls an LLM iteratively and executes tool calls.",
      "language": "rust"
    }
  ]
}

Deep Research

Spawns three researcher sub-agents in parallel, then aggregates their findings into a structured decision. Requires BRAVE_API_KEY for web search.

make use_case name=deep-research args="What constitutes a good life?"

Output:

{
  "title": "What Constitutes a Good Life: A Multi-Perspective Analysis",
  "research": "A good life emerges from the convergence of philosophical wisdom, scientific research, and cultural understanding. Key elements include meaningful relationships and social connections, a sense of purpose and personal growth, physical and mental well-being, contributing to something beyond oneself, and living in accordance with personal values. While cultural contexts vary, common themes across traditions emphasize virtue, balance, gratitude, and the cultivation of both inner fulfillment and positive impact on others."
}

Model Pricing Tracker

Spawns a model checker and pricing researcher in parallel to gather current model pricing from provider websites, then outputs structured JSON.

make use_case name=model-pricing-tracker

Output:

{
  "models": [
    {
      "input_per_million": 3.0,
      "model_id": "claude-sonnet-4-20250514",
      "output_per_million": 15.0,
      "provider": "anthropic"
    },
    {
      "input_per_million": 1.0,
      "model_id": "claude-haiku-4-5-20251001",
      "output_per_million": 5.0,
      "provider": "anthropic"
    }
  ]
}

API

An agent is configured with a provider, model, tools, and prompt. Running it returns an output with the response and statistics. Events are emitted during execution for streaming and observability.

LlmProvider

Providers for Anthropic, OpenAI-compatible, Mistral, and LiteLLM. Each owns a reqwest::Client for connection pooling and SSE streaming.

use agentwerk::{AnthropicProvider, OpenAiProvider};

let provider = AnthropicProvider::new(key);
let provider = OpenAiProvider::new(key);

// share a connection pool
let client = reqwest::Client::new();
let provider = AnthropicProvider::with_client(key, client);

Agent

Configures the agent's identity, tools, provider, and runtime options.

use agentwerk::Agent;

let output = Agent::new()
    .identity_prompt("You are a helpful assistant.")
    .instruction_prompt("What does src/main.rs do?")
    .model("claude-sonnet-4-20250514")
    .tool(ReadFileTool)
    .provider(provider)
    .run()
    .await?;

Configure a template once, clone it for each task, and override the fields that vary per run. The AgentPool example below shows this.

Prompting

Method File variant Purpose
identity_prompt identity_prompt_file Persistent identity of the agent
instruction_prompt instruction_prompt_file Task for the current run
context_prompt context_prompt_file Additional context appended after environment metadata (working directory, platform, OS version, date)
behavior_prompt behavior_prompt_file Override the default behavioral directives (DEFAULT_BEHAVIOR_PROMPT) 
Agent::new()
    .identity_prompt_file("prompts/identity.md")
    .instruction_prompt("Summarize the project.")
    .behavior_prompt_file("prompts/behavior.md")

Use {key} placeholders in the identity prompt and fill them with template_variable:

Agent::new()
    .identity_prompt("You are {role}. Respond in {language}.")
    .template_variable("role", json!("a code reviewer"))
    .template_variable("language", json!("German"))

Sub-agents

An agent can register other agents as sub-agents. The LLM can then call them by name.

let researcher_base = Agent::new()
    .model("claude-haiku-4-5-20251001")
    .identity_prompt("Research this topic.")
    .tool(brave_search_tool())
    .max_turns(3);

let r1 = researcher_base.clone().name("researcher_1");
let r2 = researcher_base.clone().name("researcher_2");

let output = Agent::new()
    .name("orchestrator")
    .identity_prompt("Coordinate research.")
    .sub_agents([r1, r2])

Registered sub-agents are available to the LLM by name. The LLM can also spawn ad-hoc agents at call time, supplying the prompt for that spawn.

Inheritance

A sub-agent is just an Agent; configure it with the normal builder methods. Three rules are specific to running as a sub-agent:

Behavior Fields
Inherited provider, model, working_directory, event_handler, cancel_signal
Shared command_queue, session_store
Own behavior_prompt, context_prompt, tools, output_schema
Per-spawn instruction_prompt (required), model, identity_prompt, max_turns, max_tokens, max_schema_retries, max_request_retries, request_retry_backoff_ms

Guardrails

Per-agent limits for agentic execution. Omit a setter to use the default (most default to "no limit"). When the parent's LLM spawns a sub-agent, it can override any of these for a single spawn.

Method Default What it does
.max_turns(10) no limit Stop after N agentic loop iterations
.max_tokens(4096) provider default Cap output tokens per LLM request
.max_schema_retries(3) 10 Retry structured output compliance
.max_request_retries(5) 3 Retry on transient API errors (429, 529, 5xx)
.request_retry_backoff_ms(2000) 10,000 Base delay for exponential backoff (ms * 2^attempt)

To abort from outside the agent, use .cancel_signal(signal) — see Inheritance for how it propagates across sub-agents.

AgentPool

Executes agents concurrently with a configurable concurrency limit. Each agent is configured independently with its own provider, prompts, and tools. Submit work with spawn() and collect results with next() or drain().

Result strategy:

  • CompletionOrder (default) — results are returned as each agent finishes.
  • SpawnOrder — results are returned in the order agents were spawned.
use agentwerk::{Agent, AgentPool, PoolStrategy, ReadFileTool};

let template = Agent::new()
    .model("claude-haiku-4-5-20251001")
    .tool(ReadFileTool);

let pool = AgentPool::new()
    .batch_size(10)
    .ordering(PoolStrategy::SpawnOrder);

for doc in ["document A", "document B"] {
    pool.spawn(
        template
            .clone()
            .provider(provider.clone())
            .instruction_prompt(format!("Summarize {doc}"))
    )
    .await;
}

let results = pool.drain().await; // Vec<(JobId, Result<AgentOutput>)>

spawn() can be called after the pool has started processing. If the pool is at capacity, it waits for a free slot.

Events

Emitted via Agent.event_handler() during execution. Each Event carries an agent_name: String plus an EventKind variant.

use agentwerk::{Event, EventKind};

let handler = Arc::new(|event: Event| match &event.kind {
    EventKind::ToolCallStart { tool_name, .. } => eprintln!("[{}] {}", event.agent_name, tool_name),
    EventKind::AgentEnd { turns } => eprintln!("[{}] done in {} turns", event.agent_name, turns),
    _ => {}
});
Kind Description
Agent AgentStart { description } Agent begins execution. description is set when the LLM spawned this agent as a sub-agent (the label it gave), and None for the top-level run.
AgentEnd { turns } Agent finishes with turn count
TurnStart / TurnEnd Turn boundaries
LLM Provider RequestStart / RequestEnd LLM request lifecycle
ResponseTextChunk Streamed text token
TokenUsage Token counts for a request
Tool Usage ToolCallStart / ToolCallEnd Tool execution lifecycle

Tools

Tools are functions the agent can call. Implement the Tool trait or use ToolBuilder for closures.

use agentwerk::{ToolBuilder, ToolResult};

let tool = ToolBuilder::new("greet", "Say hello")
    .schema(json!({...}))
    .read_only(true)
    .handler(|input, ctx| Box::pin(async move {
        Ok(ToolResult::success("Hello!"))
    }))
    .build();

Mark a tool as .read_only(true) when it has no side effects. When the LLM calls several tools in a single response, read-only calls run in parallel; everything else runs serially in order. Default is false.

Built-in tools:

Tool Description
File ReadFileTool Read a file with line numbers, offset, and limit
WriteFileTool Create or overwrite a file
EditFileTool Find-and-replace in a file
Search GlobTool Find files by pattern (e.g., **/*.rs)
GrepTool Search file contents by substring
ListDirectoryTool List directory entries with type and size
Shell BashTool::unrestricted() Execute any shell command
BashTool::new(name, pattern) Execute shell commands matching a glob pattern
Web WebFetchTool Fetch a URL and return its content as text
Utility SpawnAgentTool Delegate work to a sub-agent
TaskTool Persistent task management (create, update, list, get)
ToolSearchTool Discover available tools by keyword

AgentOutput

The result of running an agent.

output.response_raw            // free-form LLM text
output.response                // validated JSON if output_schema was set

output.statistics.input_tokens // total input tokens
output.statistics.output_tokens// total output tokens
output.statistics.requests     // number of LLM calls
output.statistics.tool_calls   // number of tool executions
output.statistics.turns        // number of agentic turns

With an output schema, the agent returns validated JSON:

let output = Agent::new()
    .output_schema(json!({
        "type": "object",
        "properties": { "category": { "type": "string" } },
        "required": ["category"]
    }))
    .max_schema_retries(3)  // retry if agent doesn't comply (default: 3)

    .run().await?;

output.response.unwrap()["category"]  // "billing"

Or load the schema from a file:

let output = Agent::new()
    .output_schema_file("schemas/category.json")
    .run().await?;

LLM Request Composition

Each LLM request is assembled from four parts. Fields are listed in the order they appear in the request.

Part Type Parameters Description
model String model() The LLM model that processes the request
max_tokens Number max_tokens() The maximum number of tokens the model can output
tool_choice ToolChoice output_schema() A constraint that forces the model to call a specific tool
system_prompt String identity_prompt()behavior_prompt() Persistent instructions that define who the agent is and how it behaves
message Message[] context_prompt()instruction_prompt() The conversation history between user and assistant, starting with metadata, context, and the task
tools ToolDefinition[] tool() The functions the model can call during execution

Development

Building and testing

make                # build (warnings are errors)
make test           # unit tests
make fmt            # format code
make clean          # remove build artifacts
make update         # update dependencies

Integration tests

Consider configuring your LLM provider (see Environment).

make test_integration                     # run all
make test_integration name=bash_usage     # run one

Use cases

make use_case                                              # list available
make use_case name=project-scanner -- ./                   # run one
make use_case name=deep-research args="What is a good life?"  # with arguments

Publishing

make bump                  # bump patch version
make bump part=minor       # bump minor version
make publish               # publish to crates.io (runs tests first)

LiteLLM proxy

Start a local LiteLLM proxy on port 4000 that forwards to a provider. Requires Docker.

make litellm                       # default: anthropic
make litellm LITELLM_PROVIDER=openai       # use OpenAI
make litellm LITELLM_PROVIDER=mistral      # use Mistral

Environment

Use cases and integration tests pick up the LLM provider from these environment variables:

Variable Description
LITELLM_PROVIDER Explicit provider selection (anthropic, mistral, openai, litellm). Skips auto-detection

Anthropic

Variable Description
ANTHROPIC_API_KEY API key (required)
ANTHROPIC_BASE_URL API URL (default: https://api.anthropic.com)
ANTHROPIC_MODEL Model (default: claude-sonnet-4-20250514)

Mistral

Variable Description
MISTRAL_API_KEY API key (required)
MISTRAL_BASE_URL API URL (default: https://api.mistral.ai)
MISTRAL_MODEL Model (default: mistral-medium-2508)

OpenAI

Variable Description
OPENAI_API_KEY API key (required)
OPENAI_BASE_URL API URL (default: https://api.openai.com)
OPENAI_MODEL Model (default: gpt-4o)

LiteLLM proxy

Variable Description
LITELLM_BASE_URL Proxy URL (default: http://localhost:4000)
LITELLM_API_KEY Auth key (optional)
LITELLM_MODEL Model (default: claude-sonnet-4-20250514)
LITELLM_PROVIDER LLM provider (default: anthropic, options: anthropic, mistral, openai)