Anthropic Banned OpenClaw Use, Then Shipped A Half-Built Replacement
Four days after blocking subscription use inside third-party harnesses, Anthropic shipped Managed Agents — their own cloud-hosted agent product. It's genuinely useful for newbies. For anyone already in Claude Code, it's missing the three features that matter.
The sequence of moves is hard to miss. Four days ago Anthropic announced you can no longer use your Claude subscription inside third-party harnesses like OpenClaw — overnight, the cheap-agent crowd had to either go local or start paying API rates. Two days ago they teased Claude Mythos, a model they say is so capable they won't release it. Today they shipped Managed Agents, their own cloud-hosted agent product, with the headline "get to production 10x faster."
Nate Herk spent three hours stress-testing the launch and came out with a useful read: this is great for someone who has never built an agent, and disappointing for anyone who already has.
The half that's good is genuinely good. The half that's missing is exactly the half that makes agent platforms feel alive.
What Managed Agents actually is
A hosted environment for running agents on Anthropic's own infrastructure. Instead of standing up containers, securing API keys, and gluing together MCP servers yourself, you describe an agent in natural language inside the Claude console and Anthropic does the wiring. The five-step setup:
- Create agent — describe what you want; Claude auto-generates the system prompt, model, MCP server list, and tools
- Configure environment — a Claude-hosted container with pre-installed packages and networking rules
- Start session — spins up the live runtime
- Connect MCPs via OAuth-style sign-in — credentials land in a vault you can share across teams
- Test — run the agent with a prompt and watch each step
Notion has a public integration where teams drag tasks to a different status column and a Managed Agent picks up the work. That's the use case Anthropic is leading with. You don't need a Claude Code subscription to use it — just an API key with $5 of credit.
Pricing
$0.08/hour per active session (idle environments are free) + standard API token costs.
If you're running a single always-on agent at $0.08/hour × 24 × 30, that's ~$58/month just for the lights-on cost. Token usage on top of that.
What's missing — and why it matters
Three features were teased but not shipped:
| Feature | Why it matters |
|---|---|
| Outcomes — define success criteria, agent self-evaluates and iterates | This is the difference between an agent that drifts off-task and one that gets things actually done. Karpathy's outcome-grounded research pattern, but baked in. |
| Multi-agent orchestration — coordinator agent delegates subtasks via a callable-agents tool | The agent-swarm pattern that's currently DIY across every harness. Native support would matter. |
| Persistent memory across sessions | Right now every wake-up is stateless. Without persistent memory, agents can't actually "work" over time — they restart from scratch each invocation. |
And one feature that wasn't even teased: scheduled cron triggers. Nate's killer point: "You're telling me Anthropic has a model so good they can't release it, but they can't give us scheduled crons in the cloud?"
Managed Agents can only be triggered by an inbound API call. So if you want the agent to wake up every 30 minutes and check ClickUp for new tasks, you have to glue that schedule together externally — trigger.dev, n8n, anything that can make HTTP calls. The thing Claude calls "build and deploy 10× faster" still requires you to bring your own clock.
The honest take
Nate's framing is right: this is who Managed Agents is for and who it isn't.
For: non-developers, marketers, ops people, founders who've been intimidated by Claude Code, anyone building their first three agents. The OAuth-style MCP connections, the natural-language agent builder, the credential vault — these are real onboarding wins.
Not for: anyone who's already running Claude Code workflows. If you've built one Claude Code project, you can do everything Managed Agents does (and more) with the Claude Agent SDK + trigger.dev for scheduling. Cheaper. More flexible. Already has cron, multi-agent patterns, and durable state.
The deeper read: this is Anthropic catching up to where the third-party harness ecosystem already lives. OpenClaw's killer features — heartbeats (cron) and Telegram chat — are still missing from Managed Agents. That's two years of harness developers learning what makes an agent feel alive, and Anthropic launched without either one.
The CLI is the actually interesting part
The one place Managed Agents shines for power users is the anthropic CLI inside Claude Code itself. You can build a Managed Agent from inside a Claude Code project that already has your full business context loaded — meaning the system prompt the Managed Agent ships with is way more robust than the one the web console builds.
Nate's example: he created a Managed Agent inside an existing Claude Code project that already knew his quarterly goals and current initiatives. The resulting system prompt was several paragraphs of contextualized direction, not the two-line stub the web flow produces.
If you're going to use Managed Agents seriously, build them from Claude Code, not the web console. And watch the credentials — the CLI defaults to putting API keys in the system prompt, which gets stored in Anthropic's logs. Build inside the web environment if you want MCPs to be the default credential path.
What to actually do this week
- If you've never built an agent, spin up one Managed Agent for a single repeatable task — competitor research, weekly summary, lead enrichment. The OAuth MCP flow is the easiest agent-deploy experience currently shipping.
- If you've been using OpenClaw for its cron + Telegram features, don't switch yet. The features that made OpenClaw worth running aren't here. Wait for the outcomes / multi-agent / memory updates.
- If you're a Claude Code power user, the only Managed Agents use case worth your time is delegating sub-tasks to a separately-hosted agent that you call via API from Claude Code. That keeps your main context lean and lets sub-agents run on a cheaper model (Haiku/Sonnet) without bloating your local session.
- Apply for early access to outcomes, multi-agent orchestration, and persistent memory. Those are the features that will actually change whether Managed Agents matter — when they ship, the product becomes worth a second look.
The Bottom Line
Anthropic launched Managed Agents the same week they killed OpenClaw subscription support — the timing isn't coincidence. The pitch is right (lower the floor for agent-building, ship to production faster) but the floor is too low: they shipped without cron, without persistent memory, without multi-agent orchestration. The third-party harness ecosystem already solved those problems years ago. Until Managed Agents catches up to where OpenClaw and trigger.dev already are, it's a great onboarding ramp and a redundant tool for anyone past their first agent. Watch the early-access features. That's where this product gets interesting.