Managed Agents vs Agent SDK vs Cloud Tasks — which harness?

Anthropic now has three ways to run autonomous agents. Each has different pricing, capabilities, and trade-offs. Here’s the comparison nobody wrote.

The three options

	Managed Agents (API)	Agent SDK (Library)	Cloud Scheduled Tasks
Runs on	Anthropic cloud	Your infrastructure	Anthropic cloud
Cost model	$0.08/hr + API tokens	API tokens only	Claude subscription (included)
Model selection	Per-agent	Per-agent	Session-level (one model)
Multi-agent	Yes (research preview)	Yes (Agent tool)	Yes (subagents)
Tools	Bash, Read, Write, Edit, Glob, Grep, WebSearch, WebFetch	Same	Same
MCP servers	Supported	Supported	Connectors only
Auth	API key	API key	Subscription
Max runtime	Hours	Unlimited (your infra)	~30 min per session
Best for	Production SaaS, long-running tasks	CI/CD, custom apps, multi-model routing	Daily automation, personal workflows

Cost math: a real example

Scenario: A daily agent that checks 40 sources, writes 5 articles, and updates tracking data. Takes ~20 minutes of active execution, uses ~50K input + 20K output tokens with Sonnet.

Approach	Cost per run	Monthly (30 runs)
Managed Agents	$0.027 runtime + ~$0.45 tokens = $0.48	$14.40
Agent SDK (Sonnet)	~$0.45 tokens = $0.45	$13.50
Cloud Tasks (Max plan)	Included in $100/mo subscription = $0 marginal	$0 (but $100/mo fixed)

If you run <3 agents: Cloud Tasks on your subscription is cheapest. If you run 10+ specialized agents: Agent SDK gives you multi-model routing (use Haiku at $1/M for data gathering, Opus at $5/M for synthesis) which can cut token costs 60%. If you need production reliability: Managed Agents handle infrastructure, monitoring, and SSE streaming.

Managed Agents: the API

from anthropic import Anthropic

client = Anthropic()

# Create a reusable agent definition
agent = client.beta.agents.create(
    name="Daily Research Agent",
    model="claude-sonnet-4-6",
    system="You are a research agent that finds and analyzes AI developments.",
    tools=[{"type": "agent_toolset_20260401"}],
)

# Create an environment with network access
env = client.beta.environments.create(
    name="research-env",
    config={"type": "cloud", "networking": {"type": "unrestricted"}},
)

# Start a session
session = client.beta.sessions.create(
    agent=agent.id, environment_id=env.id, title="Daily research"
)

# Stream results
with client.beta.sessions.events.stream(session.id) as stream:
    client.beta.sessions.events.send(session.id, events=[
        {"type": "user.message", "content": [{"type": "text", "text": "Find today's AI news"}]}
    ])
    for event in stream:
        if event.type == "agent.message":
            print(event.content[0].text)
        elif event.type == "session.status_idle":
            break

All endpoints require the managed-agents-2026-04-01 beta header. Multi-agent coordination is in research preview.

Decision framework

Start with Cloud Tasks if you’re on a Claude subscription and want zero-cost automation. It’s the fastest path and works for daily workflows.

Move to Agent SDK when you need multi-model routing (Haiku for cheap tasks, Opus for quality), CI/CD integration, or custom orchestration logic.

Use Managed Agents when you’re building a product that needs hosted agent execution, SSE streaming to a frontend, or Anthropic-managed infrastructure.