Agenter

🚀 Why Agenter?

The missing abstraction layer for embedding coding agents in your software.

Other code agent wrapper SDKs let you swap the underlying LLM backend (e.g. Claude vs. GPT-4) - Agenter lets you swap the entire Agent Runtime (e.g. Claude Code vs. Codex vs. OpenHands or custom agents).

Why does this matter? Because the LLM "brain" is only half the battle. The "body" (tools, loops, environment, workflows, and graphs) is where the real engineering differentiation lives. While Anthropic has optimized Claude Code for enterprise workflows and OpenAI has fine-tuned Codex for code generation, Agenter does not tie you to any vendor.

Agenter is a universal interface for these specialized agent runtimes. It solves the "N x M" integration problem between your apps and a fragmented agent landscape. Agenter offers

1. Runtime Agnosticism: Don't lock your platform into one vendor's agent architecture and its arbitrary conventions. Switch from a custom loop to Claude Code (or future runtimes) with a simple config change.
2. Unified Governance: Apply the same budget limits (cost or tokens) and safety policies across any underlying coding agent regardless of how it executes code.
3. Workflow Portability: Write your LangGraph or PydanticAI logic once. Run it on the best coding engine available today or tomorrow.

At the core of Agenter is AutonomousCodingAgent — a facade that handles all heavy backend selection, tool execution, validation loops, budget enforcement, event streaming, and other concerns. You give it a prompt and a directory, and it writes working code for you.

🏗️ Architecture

The SDK is built on a robust CodingBackend abstraction protocol that decouples agent logic from underlying backend providers.

┌─────────────────────────────────────────────────────────────┐
│                        User Applications                    │
├─────────────────────────────┬───────────────────────────────┤
│      LangGraph Adapter      │      PydanticAI Adapter       │
└─────────────────────────────┴───────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│               AutonomousCodingAgent (Facade)                │
│       (Unified API, Configuration, Event Streaming)         │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                       CodingSession                         │
│   (Iteration Loop, Budget Enforcement, Error Recovery)      │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌───────────────────────────────────────────────────────────────────────────────────────────────────┐
│                                CodingBackend Protocol (Abstract)                                  │
├───────────────────────┬───────────────────────┬─────────────────────┬─────────────────┬───────────┤
│  AnthropicSDKBackend  │   ClaudeCodeBackend   │    CodexBackend     │ OpenHandsBackend│ ACPBackend│
│  (Custom Tool Loop)   │  (Claude Code SDK)    │   (OpenAI Models)   │  (Any Model)    │ (ACP CLI) │
└───────────────────────┴───────────────────────┴─────────────────────┴─────────────────┴───────────┘

📦 Installation

# Default installation (anthropic-sdk backend with Anthropic API or AWS Bedrock)
pip install agenter

# Installation with Claude Code support (claude-code-sdk)
pip install agenter[claude-code]

# Installation with Codex support (OpenAI models)
pip install agenter[codex]

# Installation with OpenHands support (any model via litellm)
pip install agenter[openhands]

# Installation with ACP support (Agent Client Protocol agents)
pip install agenter[acp]

# Installation with specific framework adapters
pip install agenter[langgraph,pydantic-ai]

# Installation with security scanning (Bandit)
pip install agenter[security]

📋 Backend Requirements

The default anthropic-sdk backend works as a pure Python library - just set your API key and go. Other backends wrap external CLIs/services and require additional setup.

⚡ Quick Start

The SDK manages the core coding loop for you. Just provide a prompt and a directory, and you are good to go.

import asyncio
from agenter import AutonomousCodingAgent, CodingRequest, Verbosity

async def main():
    # 1. Initialize the coding agent (defaults to anthropic-sdk backend)
    agent = AutonomousCodingAgent(
        model="claude-sonnet-4-20250514",
    )

    # 2. Execute a task with full observability
    print("🤖 Coding agent starting...")
    result = await agent.execute(
        CodingRequest(
            prompt="Create a FastAPI app with a health check endpoint and a test file for it.",
            cwd="./workspace",
            allowed_write_paths=["*.py"]  # Security: Only allow Python files to be written
        ),
        verbosity=Verbosity.NORMAL  # Shows progress and tool calls in real time
    )

    # 3. Check results
    if result.status == "completed":
        print(f"✅ Success! Modified {len(result.files)} files.")
        print(f"💰 Cost: ${result.total_cost_usd:.4f}")
    else:
        print(f"❌ Failed: {result.summary}")

if __name__ == "__main__":
    asyncio.run(main())

🛡️ Unified Sandbox Mode

Sandboxing prevents the coding agent from modifying files outside the working directory, protecting your system against unintended changes.

All of the following coding agent backends default to sandbox=True for safe operation.

Sandbox Enabled (Default)

# All backends run in safe mode by default
agent = AutonomousCodingAgent(backend="anthropic-sdk")  # PathResolver enforces cwd
agent = AutonomousCodingAgent(
    backend="claude-code",
    model="claude-sonnet-4-5-20250929",
    claude_max_thinking_tokens=8192,
)  # Native OS-level sandbox
agent = AutonomousCodingAgent(
    backend="codex",
    model="gpt-5.4",
    codex_reasoning_effort="high",
)  # Workspace-write mode
agent = AutonomousCodingAgent(
    backend="acp",
    acp_command="codex-acp",
    acp_args=["-c", 'model="gpt-5.4"', "-c", 'model_reasoning_effort="high"'],
    acp_permission_policy="allow",
)

ACP prompts run with Agenter's autonomous backend contract by default, so interactive agents do not pause for routine implementation confirmation. Set acp_autonomous=False if you want the raw ACP agent behavior.

Persistent ACP Follow-ups

Use open_session() when an ACP coding agent should retain its own conversation state across multiple prompts instead of receiving one large generated program:

from agenter import AutonomousCodingAgent, Budget

agent = AutonomousCodingAgent(
    backend="acp",
    acp_command="codex-acp",
    acp_permission_policy="allow",
)
session = await agent.open_session(
    cwd="./workspace",
    session_budget=Budget(max_iterations=40, max_tokens=200_000),
)
try:
    first = await session.execute("Inspect the GLB and implement the extraction script.")
    followup = await session.execute("Run it, inspect the render, and fix the wall contacts.")
    print(session.session_id, followup.request_index, session.modified_files().paths())
finally:
    await session.close()

Follow-ups are serialized onto one ACP subprocess and session/prompt session. Each result reports request-local files and usage plus stable session_id, request_index, and cumulative session totals. session.modified_files() and session.usage() expose cumulative state. Active prompts can be cancelled with session.cancel() without discarding the session. Pass resume_session_id to open_session() when the ACP agent advertises session/resume or legacy session/load support. Check result.usage_reported (or session.usage().reported) because some ACP adapters do not publish token usage.

⚠️ OpenHands Warning: The OpenHands backend requires sandbox=False and has no filesystem isolation. It can read and write anywhere on your system. Use it with caution and only in trusted environments.

⚠️ ACP Warning: The ACP backend launches an external agent process. Agenter can observe changed files after the run, but filesystem isolation depends on the ACP agent and its own flags.

Sandbox Disabled (Full File System Access)

# Disable sandbox for full file system access
agent = AutonomousCodingAgent(backend="claude-code", sandbox=False)

🧩 Agentic Framework Integrations

Don't reinvent the wheel. Use our pre-built adapters to supercharge your agentic graphs.

Adapters subclass framework base classes to include LangSmith/Logfire tracing and all framework features automatically.

LangGraph Adapter

create_coding_node() returns a RunnableLambda (includes LangSmith tracing).

from agenter.adapters.langgraph import create_coding_node

node = create_coding_node(cwd="./workspace", backend="claude-code")
result = await node.ainvoke({"prompt": "Create a hello world script"})

print(result["coding_result"]["summary"])

PydanticAI Adapter

CodingAgent subclasses pydantic_ai.Agent (includes Logfire tracing).

from agenter.adapters.pydantic_ai import CodingAgent

agent = CodingAgent(cwd="./workspace", backend="codex")
result = await agent.run("Add input validation")

print(result.summary)

🎯 Structured Outputs

Get typed and fully validated outputs from your coding tasks using Pydantic models.

Basic Usage

from pydantic import BaseModel
from agenter import AutonomousCodingAgent, CodingRequest

class AnalysisResult(BaseModel):
    summary: str
    issues_found: list[str]
    confidence: float

agent = AutonomousCodingAgent()
result = await agent.execute(
    CodingRequest(
        prompt="Analyze the security of auth.py",
        cwd="./workspace",
        output_type=AnalysisResult,  # Enables structured outputs
    )
)

# result.output is typed as AnalysisResult
print(result.output.summary)
print(result.output.confidence)

🔍 Observability

The SDK provides rich, structured event streaming for UIs or logs.

async for event in agent.stream_execute(request):
    if event.type == "backend_message":
        print(f"🤖 AI: {event.data['content']}")
    elif event.type == "validation_start":
        print("🔍 Validating code...")
    elif event.type == "iteration_end":
        print(f"🔄 Iteration {event.data['iteration']} finished.")

📚 Documentation

• Quickstart Notebook — Get started with an interactive tutorial.
• Architecture — Dive deeper into the SDK's layered design, backend protocol, event system, and functionality.
• Naming Conventions — Supplementary material about coding style and naming patterns.

📜 License

Copyright 2025 Moonsong Labs

Licensed under the Apache License, Version 2.0. See LICENSE for details.