Agenkit

Build production-ready AI agent systems.

Agenkit is a lightweight, cross-language toolkit for building distributed AI agents that scale from prototype to production without rewriting your code.

Why Agenkit?

The Problem

Building production AI agent systems is hard:

• Reliability: LLMs fail unpredictably - you need circuit breakers, retries, and timeouts
• Scale: Prototypes work locally but break in production when you need distributed deployment
• Observability: Understanding what went wrong requires distributed tracing across services
• Language Lock-in: Python is great for prototyping, but you need Go/Rust for performance
• Integration: Every agent framework has its own incompatible abstractions

The Solution

Agenkit provides the production infrastructure you need:

Prototype → Production
┌─────────────────────────────────────────────┐
│  Your Agents (Python for development)      │
└──────────────┬──────────────────────────────┘
               │
               │ Same Interface
               ↓
┌─────────────────────────────────────────────┐
│  Agenkit Toolkit                            │
│  • Automatic retries & circuit breakers    │
│  • Distributed tracing & metrics           │
│  • Cross-language support (Python ↔ Go)    │
│  • Multiple transports (HTTP/gRPC/WS)      │
└─────────────────────────────────────────────┘

Key Insight: Write your agents once in Python. Deploy them in Go for up to 18x lower framework/transport overhead in our benchmarks (see Performance below) — the LLM call itself still dominates wall-clock time. Same interface, zero rewrites.

Quick Start

30 Second Example

Create a simple agent:

from agenkit import Agent, Message

class MyAgent(Agent):
    @property
    def name(self) -> str:
        return "my-agent"

    async def process(self, message: Message) -> Message:
        return Message(
            role="agent",
            content=f"Processed: {message.content}"
        )

# Use it
agent = MyAgent()
response = await agent.process(Message(role="user", content="Hello!"))

Add Production Features in 3 Lines

from agenkit.middleware import RetryConfig, RetryDecorator, CircuitBreakerDecorator

# Wrap with resilience
production_agent = RetryDecorator(
    CircuitBreakerDecorator(agent),
    RetryConfig(max_retries=3),
)

# Now handles failures automatically
response = await production_agent.process(message)

Deploy Anywhere

# Run locally
python my_agent.py

# Deploy with Docker
docker-compose up

# Scale with Kubernetes
kubectl apply -f deploy/kubernetes/

Core Features

🎯 Simple, Minimal Interface

class Agent:
    name: str                          # Unique identifier
    async def process(msg) -> Message  # Process messages

That's it. Everything else is optional.

🔄 Production Middleware (Add What You Need)

from agenkit.middleware import (
    CircuitBreakerConfig,
    CircuitBreakerDecorator,
    RetryConfig,
    RetryDecorator,
    TimeoutConfig,
    TimeoutDecorator,
)

# Start simple
agent = MyAgent()

# Add retry logic
agent = RetryDecorator(agent, RetryConfig(max_retries=3))

# Add circuit breaker
agent = CircuitBreakerDecorator(agent, CircuitBreakerConfig(failure_threshold=5))

# Add timeouts (timeout_ms for clarity)
agent = TimeoutDecorator(agent, TimeoutConfig(timeout_ms=30000))

# Stack as many as you need

Every middleware is <100 lines. Easy to understand, modify, or replace.

🌐 Cross-Language Support

Write once. Deploy anywhere. Nine language implementations (Python is the reference; all nine share the same Agent/Message/Tool core and most of the 18 core patterns — C#, Java, and Scala are missing AgentsAsTools; see Status for the exact per-language spec-conformance breakdown):

# Python - Prototype quickly with ML ecosystem
class MyAgent(Agent):
    async def process(self, message):
        return process_with_python_libs(message)

// TypeScript - Full-stack with browser support
class MyAgent implements Agent {
    async process(message: Message): Promise<Message> {
        return processWithTypeScriptLibs(message);
    }
}

// Go - Production scale (up to 18x lower transport overhead vs Python in our benchmarks)
type MyAgent struct{}

func (a *MyAgent) Process(ctx context.Context, msg *Message) (*Message, error) {
    return processWithGoLibs(msg)
}

// C++ - Maximum performance with zero-overhead abstractions
class MyAgent : public Agent {
public:
    Message process(const Message& msg) override {
        return process_with_cpp_libs(msg);
    }
};

// Rust - Memory safety + performance (up to 20x lower transport overhead vs Python in our benchmarks)
struct MyAgent;

impl Agent for MyAgent {
    async fn process(&self, message: Message) -> Result<Message, AgentError> {
        process_with_rust_libs(message).await
    }
}

// Zig - Systems programming with safety (up to 22x lower transport overhead vs Python in our benchmarks)
const MyAgent = struct {
    pub fn process(self: *MyAgent, message: Message) !Message {
        return processWithZigLibs(message);
    }
};

Same interface across all languages. Python, Go, and Rust are the most complete; C#, Java, and Scala are newer and still filling in advanced subsystems (skills, some adapters). Choose the right tool for each service.

📊 Full Observability

from agenkit.observability import TracingMiddleware, init_tracing

# Enable tracing
init_tracing("my-service")

# Wrap agent
traced_agent = TracingMiddleware(agent)

# Every operation now traced in Jaeger

View traces: http://localhost:16686

🔍 Agent Introspection

# Inspect agent state during debugging or production
result = agent.introspect()

print(f"Agent: {result.agent_name}")
print(f"Capabilities: {result.capabilities}")
print(f"Internal state: {result.internal_state}")
print(f"Memory: {result.memory_state}")

# Use for monitoring
def check_agent_health(agent):
    result = agent.introspect()
    return result.internal_state.get("error_count", 0) < 10

# Use for testing
def test_agent_state():
    result1 = agent.introspect()
    await agent.process(message)
    result2 = agent.introspect()
    assert result2.internal_state["msg_count"] > result1.internal_state["msg_count"]

Introspection ≠ Reflection:

• Introspection (this feature): Examines current state ("What do I know?")
• Reflection (pattern): Analyzes past performance ("How did I do?")

🚀 Multiple Transports

from agenkit.adapters.python import GRPCServer, LocalAgent
from agenkit.adapters.python.http_server import HTTPAgentServer

# Same agent, different transports
await HTTPAgentServer(agent, port=8080).start()               # REST API
await GRPCServer(agent, "localhost:50051").start()            # High performance
await LocalAgent(agent, endpoint="ws://127.0.0.1:8765").start()  # Bidirectional streaming

Choose the right protocol for each use case.

When Should You Use Agenkit?

✅ Perfect For:

• Building production AI agent systems that need to scale
• Teams that prototype in Python but deploy in Go
• Systems that need resilience (retries, circuit breakers, timeouts)
• Distributed agent systems that need observability
• Integrating multiple AI agent frameworks

❌ Not For:

• Simple one-off scripts (too much overhead)
• Embedded systems (requires async runtime)
• Real-time systems (<1ms latency requirements)

Installation

Core Installation

# Python
pip install agenkit

# Go
go get github.com/scttfrdmn/agenkit-go

# TypeScript/Node.js
npm install @agenkit/core

# C++
# Clone and build (CMake required)
git clone https://github.com/scttfrdmn/agenkit.git
cd agenkit/agenkit-cpp && mkdir build && cd build && cmake .. && make

# Rust
cargo add agenkit

# Zig
# Clone and build (Zig 0.12+ required)
git clone https://github.com/scttfrdmn/agenkit.git
cd agenkit/agenkit-zig && zig build

Python: Install with Specific LLM Providers

# Install only what you need
pip install agenkit[openai]          # OpenAI (GPT-4, GPT-3.5)
pip install agenkit[anthropic]       # Anthropic (Claude 3.5)
pip install agenkit[aws]             # AWS Bedrock
pip install agenkit[google]          # Google Gemini
pip install agenkit[ollama]          # Ollama (local models)

# Multiple providers
pip install agenkit[openai,anthropic]

# All providers
pip install agenkit[all-providers]

# With Redis memory backend
pip install agenkit[redis]

# Everything (for development)
pip install agenkit[all]

See INSTALLATION.md for complete installation guide.

What's Included?

Core Components

• Minimal interfaces - Agent, Message, Tool (50 lines total)
• Orchestration patterns - Sequential, Parallel, Router, Fallback
• Type safety - Full type hints (Python), compile-time checks (Go)

Production Middleware (Optional)

• Circuit Breaker - Prevent cascading failures
• Retry - Exponential backoff with jitter
• Timeout - Request deadline enforcement
• Rate Limiter - Token bucket algorithm
• Caching - LRU cache with TTL
• Batching - Request aggregation

Transport Layer (Optional)

• HTTP - REST APIs (HTTP/1.1, HTTP/2, HTTP/3)
• gRPC - High-performance binary protocol
• WebSocket - Bidirectional streaming

Observability (Optional)

• Distributed Tracing - OpenTelemetry integration
• Metrics - Prometheus endpoints
• Structured Logging - JSON logs with trace correlation

Autonomous Agent Features (Optional)

• Memory Management - Context retention with compression strategies
• Budget Tracking - Token counting and cost optimization
• Checkpointing - State persistence and recovery
• Safety - Prompt injection detection, input/output validation
• Evaluation - Quality metrics and benchmarking

Performance

Transport Overhead: <1% of total time in realistic LLM workloads

Language Performance (framework/transport overhead, not general application performance):

• Go HTTP transport: 18.5x lower latency than Python's in the same benchmark (0.055ms vs 1.02ms)
• Middleware overhead: <0.01% of request time

These figures — and the "18x"/"20x"/"22x" multipliers mentioned earlier in this README for Go/Rust/Zig — were measured in November 2025 against Python 3.14.0 and Go 1.21–1.22 on Apple Silicon (see benchmarks/BASELINES.md). Both the Go toolchain and several dependency versions have since moved, so the exact multiplier should not be read as current; the qualitative conclusion that transport/framework overhead is a small fraction of a 100–1000ms LLM call has held across every measurement to date. See COMPATIBILITY.md for the full caveat and how to regenerate current numbers.

Scale:

• Kubernetes autoscaling: 3-10 replicas based on load
• Message size scaling: 10,000x size = 190x latency (linear)

See benchmarks/BASELINES.md for detailed performance data.

Production Ready

8500+ Tests Passing

• Cross-language integration tests (Python ↔ Go ↔ C++)
• Chaos engineering tests (network failures, crashes)
• Property-based tests (invariant validation)
• Thousands of unit and integration tests across 9 languages (Python alone: 2200+; see Test Parity below for the current per-language breakdown)

Security

• Input validation
• Prompt injection detection
• RBAC and audit logging
• Non-root containers
• Dropped capabilities

Observability

• Jaeger distributed tracing
• Prometheus metrics
• Structured JSON logging
• Health check endpoints

Deployment

• Docker images
• Kubernetes manifests with HPA
• Horizontal pod autoscaling (3-10 replicas)
• Zero-downtime rolling updates

Architecture

┌─────────────────────────────────────────┐
│   Your Application (Agents + Logic)    │
└──────────────┬──────────────────────────┘
               │
┌──────────────┴──────────────────────────┐
│   Optional Middleware Layer             │
│   (Add only what you need)              │
│   • Retry     • Caching                 │
│   • Timeout   • Batching                │
│   • Circuit Breaker                     │
└──────────────┬──────────────────────────┘
               │
┌──────────────┴──────────────────────────┐
│   Optional Transport Layer              │
│   • HTTP   • gRPC   • WebSocket         │
└──────────────┬──────────────────────────┘
               │
┌──────────────┴──────────────────────────┐
│   Core Interface (Required)             │
│   • Agent  • Message  • Tool            │
└─────────────────────────────────────────┘

Design Philosophy: Start simple. Add complexity only when you need it.

Learning Path

1. Getting Started (15 min) - Choose your language and create your first agent
   • Python Guide | Go Guide | TypeScript Guide
   • Rust Guide | C++ Guide | Zig Guide
2. Agent Patterns Book (2-3 hours) - Master the 18 core patterns and advanced architectures
3. Advanced Architectures (1 hour) - Pattern compositions and multi-agent systems
4. Examples (1 hour) - Learn by doing with 27+ examples
5. Architecture (30 min) - Understand design principles
6. Migration Guides (30 min per language) - Migrate between languages
7. Production Deployment (1 hour) - Docker + Kubernetes

Total time investment: ~6 hours from zero to production deployment with pattern mastery.

Documentation

Getting Started (Language-Specific)

• Python Guide - 15-30 min to first agent
• Go Guide - Idiomatic Go patterns
• TypeScript Guide - Modern JavaScript/TypeScript
• Rust Guide - Memory-safe agents
• C++ Guide - High-performance systems
• Zig Guide - Systems programming with safety

Patterns & Architecture

• Agent Patterns Book - Comprehensive guide to 18 core patterns
• Advanced Architectures - Pattern compositions and multi-agent systems
• Architecture Principles - Design philosophy
• Streaming Patterns - Language-specific streaming approaches

Migration & Reference

• Migration Guide Index - Complete migration documentation for all 6 languages
• API Reference - Complete API documentation
• Deployment Guide - Docker and Kubernetes
• Examples - 27+ comprehensive examples

Operations & Security

• Security Policy - Vulnerability reporting and best practices
• Compatibility Matrix - Language, platform, and version support
• Testing Strategy - Test philosophy and coverage

Package-Specific Docs

• Memory Management - Context retention strategies
• Budget Tracking - Token and cost management
• Checkpointing - State persistence
• Safety & Security - Input validation
• Evaluation - Quality metrics

Examples

We provide 27+ examples covering common use cases:

Getting Started:

• Basic Agent - Your first agent in 20 lines
• Sequential Pipeline - Chain agents together
• Parallel Execution - Run agents concurrently

Production Features:

• Retry & Circuit Breaker - Handle failures
• Distributed Tracing - Debug with Jaeger
• Rate Limiting - Protect your APIs

Advanced:

• Remote Agents - Cross-process communication
• Streaming Responses - Server-sent events

Browse all examples →

Community

• Website: agenkit.dev
• GitHub: github.com/scttfrdmn/agenkit
• Issues: Report bugs or request features
• Discussions: Ask questions

Contributing

We welcome contributions! See our Contributing Guide.

Development Setup

# Clone repository
git clone https://github.com/scttfrdmn/agenkit.git
cd agenkit

# Install dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Run type checking
mypy agenkit/

License

Apache License 2.0 - See LICENSE for details.

Status

v0.89.0 — 9 language implementations, Python reference (see the root VERSION file for the current number — this line will drift again if hand-maintained)

Language Support

The shared core is the Agent/Message/Tool interface plus 18 named patterns (listed below). The "Patterns implemented" column is a spec-presence conformance score from spec-conformance.json (scripts/parity/spec_conformance.py, #909/#913) — for each of the 18 patterns named in specs/patterns/*.yaml, does a source file implementing it actually exist in that language, checked by filename/class-name (with an alias table for legitimate naming divergence, e.g. Python's memory.py/MemoryHierarchy vs. C#'s MemoryAugmentedAgent.cs) rather than by counting classes. This replaced an earlier raw class-count column (feature-manifest.json's patterns count) that answered a different, noisier question — it could not distinguish "no implementation exists" from "the implementation has an unconventional class name," and undercounted C#/Java/Scala for months due to a scanner bug (#918) before this metric existed. feature-manifest.json remains as a secondary, diagnostic class-count signal; it is no longer this table's source. The "Tests" column is the test count from the parity report; "Depth" reflects how many advanced subsystems (memory, skills, reasoning memory, full adapter set) are implemented.

Language	Patterns implemented	LLM Adapters	Tests	Depth
Python	18/18	7	2229	Reference — all subsystems
Go	18/18	7 (+vLLM, SGLang)	1341	Complete — incl. reasoning memory, skills
TypeScript	18/18	7	976	Broad — no skills / reasoning memory
Rust	18/18	6	1352	Complete — incl. skills
C++	18/18	5	1133	Broad — safety/ not yet implemented
Zig	18/18	8	671	Broad — no skills
C# (.NET)	17/18 (missing AgentsAsTools)	2 (+mock)	272	Newer — no skills
Java	17/18 (missing AgentsAsTools)	2 (+mock)	358	Newer — no skills
Scala	17/18 (missing AgentsAsTools)	mock only	363	Newest — LLM adapters are stubs

18 Core Patterns documented in the Agent Patterns Book: AgentsAsTools, Autonomous, Collaborative, Conversational, Fallback, HumanInLoop, MemoryHierarchy, MultiAgent, Orchestration, Parallel, Planning, ReAct, ReasoningWithTools, Reflection, Router, Sequential, Supervisor, Task.

Six languages (Python, Go, TypeScript, Rust, C++, Zig) implement all 18; C#, Java, and Scala implement 17 of 18 (missing AgentsAsTools).

Recent Highlights (v0.85 – v0.89)

• ✅ Defect repair (v0.89) — 82 commits fixing release-gate and CI bugs found by a fleet audit (SBOM/signing, version-declaration drift, a test gate that couldn't fail)
• ✅ Typed cross-language token Usage (v0.86–v0.87) — unified Usage struct across all 9 language cores, plus Bedrock prompt-cache token counts
• ✅ Agent Skills (v0.85) — AgentSkill, SkillRegistry, SkillEnabledAgent (Python, Go, Rust)

v0.88.0 is intentionally skipped — reserved for the observability milestone (#715). See CHANGELOG.md for the full release history.

Project Status

• ✅ Core toolkit across 9 languages; 18 shared patterns in 6 of them, 17 of 18 in C#/Java/Scala (see Status)
• ✅ MCP (Model Context Protocol) client/server in every language
• ✅ Production middleware (retry, circuit breaker, timeout, rate limiting, caching, batching)
• ✅ Multiple transports (HTTP/1.1, HTTP/2, HTTP/3, gRPC, WebSocket)
• ✅ Deployment manifests (Docker + Kubernetes with HPA)
• 🚧 Advanced-subsystem parity (skills, reasoning memory, full adapter sets) still landing in the JVM/.NET tier

---

Ready to build production AI agents? Visit agenkit.dev or get started in 5 minutes →

Test Parity

Patterns are at full parity across languages; test-count parity varies — the secondary languages have fewer tests than the Python reference. Counts are regenerated via scripts/test-parity.sh (and surfaced by the Parity Validation CI workflow). Relative to Python's 2229 tests:

Language	Tests	vs Python
Rust	1352	61%
Go	1330	60%
C++	1133	51%
TypeScript	976	44%
Zig	671	30%
Java	358	16%
Scala	363	16%
C#	272	12%

Counts are regenerated via scripts/test-parity.sh.