Agent frameworks compared

There are more agent frameworks than anyone can try, and new ones every month. The good news: underneath, they all run the same loop from What an agent is. They differ in three questions, and once you can ask those questions, choosing one stops being a matter of hype.

Who owns the loop? You, a library, or a hosted service.

What comes built in? Nothing but the loop, or persistence, multi-agent handoffs, built in file and shell tools, tracing and deployment.

How tied are you to one model provider?

This chapter writes the same small agent in five ways, then lays the wider field out in one table.

One agent, five ways#

The task: a support agent with a single tool, get_order_status, answering "Where is order 44821?". Switch between the tabs and notice how much is the same.

# The loop by hand, with the Anthropic Python SDK. Everything is visible.
import anthropic
 
client = anthropic.Anthropic()
tools = [{
    "name": "get_order_status",
    "description": "Look up the delivery status and ETA of one order by its id.",
    "input_schema": {
        "type": "object",
        "properties": {"order_id": {"type": "string"}},
        "required": ["order_id"],
    },
}]
messages = [{"role": "user", "content": "Where is order 44821?"}]
 
for _ in range(10):  # a step budget, enforced by your code
    response = client.messages.create(
        model="claude-opus-5-5", max_tokens=16000, tools=tools, messages=messages
    )
    messages.append({"role": "assistant", "content": response.content})
    if response.stop_reason != "tool_use":
        break
    results = [
        {"type": "tool_result", "tool_use_id": block.id,
         "content": orders_api.status(block.input["order_id"])}
        for block in response.content if block.type == "tool_use"
    ]
    messages.append({"role": "user", "content": results})

In every tab the same four things appear: a model, a tool with a name, a description and a schema, the loop (written, configured or hidden), and the answer at the end. The frameworks differ in what they add around that core, not in the core.

A different kind of thing: agents with a built in toolbox#

The SDKs above give you a loop and expect you to supply every tool. A newer category hands you a complete agent with tools already wired: read and edit files, run shell commands, search, manage its own context, spawn sub agents. The Claude Agent SDK is Claude Code packaged as a library; OpenAI's Codex has a similar SDK. You describe the task and the permissions, not the loop.

import { query } from "@anthropic-ai/claude-agent-sdk";
 
for await (const message of query({
  prompt: "test_checkout fails after the coupon change. Find out why and fix it.",
  options: {
    allowedTools: ["Read", "Edit", "Glob", "Grep", "Bash"],
    permissionMode: "acceptEdits",
  },
})) {
  if (message.type === "result") console.log(`Done: ${message.subtype}`);
}

That is the whole program. It is the right choice when the agent's job looks like an engineer's job (working in a repository, a filesystem, a terminal) and the wrong one for a narrow support bot that should only ever call three of your APIs.

Where they differ#

Hosted options sit one level higher: the major model providers now offer managed agent services that run the loop and a sandbox for you, which trades control for not running any infrastructure.

Checkpoint#