Dev Tools

If you've been living under a rock, every framework wants you to draw an agent graph in YAML. Behold: PrefectHQ's Marvin just slaps a decorator on a function and you're done.

The Setup

Marvin's core idea: an LLM call is just a function with a signature. Pydantic in, Pydantic out, docstring is the prompt. It plays perfectly with the Cursor + Claude Code daily workflow because the types are visible everywhere.

import marvin
from pydantic import BaseModel

class Claim(BaseModel):
    address: str
    damage_type: str
    severity: int  # 1-5

@marvin.fn
def extract_claim(transcript: str) -> Claim:
    """Pull insurance claim details from a phone call transcript."""

result = extract_claim("Caller at 12 Main St, hail through the roof, bad")
print(result)  # Claim(address='12 Main St', damage_type='hail', severity=4)

The Money Pattern

The killer feature is composability. You stack \`@marvin.fn\` and \`@marvin.classifier\` calls like regular Python — no graph, no nodes, no event bus. The whole agent reads top-to-bottom like a script.

from typing import Literal

@marvin.classifier
def route(text: str) -> Literal["billing", "support", "sales"]:
    """Route an inbound message to the right team."""

@marvin.fn
def draft_reply(text: str, team: str) -> str:
    """Draft a short reply for the given team."""

msg = "I want a refund for last month"
team = route(msg)
reply = draft_reply(msg, team)

The Catch

Marvin pulls in Prefect-flavoured dependencies and they're not small. If you're shipping a tiny serverless function, Mirascope is leaner. And the default model selection wants OpenAI — swap to Anthropic or local via \`marvin.settings\` before you ship.

The Verdict

For Pydantic-heavy codebases — which is basically every Aidxn Design backend — Marvin is the cleanest agent abstraction shipped this year. If you've been hand-rolling tool calls, give it a weekend. You won't go back to graph YAML.