Anthropic just mass-produced the agent. who's building the hands?

Yesterday Anthropic shipped Claude Managed Agents. If you haven't seen it yet — they're now hosting the full agent runtime for you. Session persistence, sandboxed execution, error recovery, tool orchestration. You define what the agent should do, they handle everything underneath.

It's a big deal. But I think most people are looking at the wrong part.

Everyone's talking about the brain. Nobody's talking about the hands.

Managed Agents gives you a brain that can reason, plan, and recover from failures across long-running sessions. It gives you bash, file I/O, web search. The infrastructure layer is genuinely impressive — $0.08/hr per agent, auto-recovery on disconnect, built-in prompt caching.

But here's what it can't do out of the box: anything domain-specific.

It can't check if your AliExpress supplier went offline. It can't search for a replacement. It can't compare SKU variant mappings between two suppliers who call the same color "Navy Blue" and "深蓝色" respectively. It can't push a product to your Shopify store.

The brain is there. The hands are missing.

That's where MCP servers come in. Managed Agents natively supports MCP — you declare your servers in the agent config, and the agent discovers and calls your tools like it would call bash or web search. No extra plumbing.

What I actually plugged in

I've been building an open-source MCP server for dropshipping automation — it connects to DSers and handles everything from product sourcing to store push. Last week I added a supplier replacement tool that searches for alternatives, scores candidates, matches SKU variants, and optionally swaps the mapping. The kind of supply chain operation that used to take a seller 2-3 hours of manual work per product.

Connecting it to Managed Agents took about 12 lines:

from claude_agent_sdk import query, ClaudeAgentOptions

async for message in query(
    prompt=(
        "Check product dp-8291 in store st-102. "
        "If the current supplier's price increased more than 20%, "
        "find a cheaper alternative and update the mapping."
    ),
    options=ClaudeAgentOptions(
        mcp_servers={
            "dsers": {
                "command": "npx",
                "args": ["-y", "@lofder/dsers-mcp-product"],
            }
        }
    ),
):
    print(message)

That's it. The agent picks up the MCP tools, figures out the right call sequence — fetch current mapping, search suppliers, score candidates, match variants, apply if confidence is high enough — and runs it autonomously. If the session disconnects, Managed Agents picks up where it left off.

Why this matters for supply chain automation

Dropshipping supply chains break constantly. Suppliers vanish, prices spike overnight, stock runs dry on your best-selling variants. Traditionally you find out when a customer complains. Or if you're diligent, you check manually every few days.

With a Managed Agent running 24/7, you can set up something like: "every 6 hours, scan my top 50 products, flag any supplier that's down or 15%+ more expensive, and auto-replace if a high-confidence alternative exists." The agent handles the scheduling loop, the MCP server handles the domain logic. You pay $0.08/hr for the runtime and whatever tokens the model consumes.

Honestly, the token cost is the part I'm still watching. My supplier replacement tool does a lot of API calls — search, detail fetch, scoring, variant matching — and each of those round-trips adds context. For a single product check it's fine. For 50 products on repeat, I haven't done the math yet. That's going to depend heavily on how well Managed Agents handles context compaction between iterations.

The interesting tradeoff nobody's mentioning

Managed Agents is opinionated about architecture. Anthropic hosts the brain, the sandbox, the session. You bring the tools. This is great if you want to move fast — you don't build an agent loop, you don't manage containers, you don't deal with state persistence.

But it also means your agent only runs on Claude. You can't swap the model. You can't self-host. You're renting the runtime from Anthropic and paying per-hour plus per-token.

For my use case that's acceptable — dropshipping sellers already pay for Shopify, DSers, ad platforms. Adding $0.08/hr for an agent that monitors their supply chain is cheap compared to losing a best-seller for 3 days because the supplier silently went 404.

But if you're building something where model portability matters, or where you need to run in your own VPC for compliance, Managed Agents isn't the answer. The MCP server still works everywhere else — Cursor, Claude Desktop, any MCP-compatible client. The Managed Agents integration is just one more deployment target.

What's still missing

A few things I ran into during setup:

• No cron/schedule primitive. You can't tell a Managed Agent "run this every 6 hours." You'd need an external scheduler (Lambda, cron job, whatever) that creates sessions on a timer. Seems like an obvious addition they'll probably ship soon.
• MCP auth in the sandbox. My MCP server uses OAuth — the user runs npx @lofder/dsers-mcp-product login once locally and the token persists. In the Managed Agents sandbox, you'd need to inject the token via environment variables or mount a credentials file. Workable, but not as smooth as the local experience.
• Observability. I can see events in the Claude Console, but I want structured logs — "agent checked 50 products, replaced 3 suppliers, skipped 2 due to low confidence." Right now I'd have to parse that from the event stream myself.

None of these are blockers. They're the kind of rough edges you expect from a beta that launched literally yesterday.

The bigger picture

MCP servers have been around for a while, but they've mostly lived in developer tools — Cursor, Claude Desktop, VS Code. Managed Agents changes the distribution model. Now an MCP server isn't just a dev tool plugin. It's a building block for autonomous business processes.

A supply chain monitoring agent. An inventory rebalancing agent. A competitive pricing agent. These are all just "a brain + the right MCP tools + a schedule." Managed Agents handles the first part. The tools are what make it useful.

If you're building MCP servers, I think this is worth paying attention to. The surface area for your tools just got a lot bigger.

---

The MCP server is open source: github.com/lofder/dsers-mcp-product. The supplier replacement tool (dsers_replace_mapping) ships in the next release. If you're running dropshipping on DSers, it works today with Cursor and Claude Desktop, and now with Managed Agents too.

Curious if anyone else has tried plugging production MCP servers into Managed Agents — what was your experience? I'm especially interested in how token costs scale for long-running sessions.