Operate Claude Managed Agents inside your security perimeter
Anthropic’s new self-hosted Claude Managed Agents lets enterprises run the entire agent runtime inside their own cloud or data center while still calling Claude wherever they like. The real move isn’t “on-prem AI” in general; it’s turning the managed agent harness into something you can drop behind your own firewalls, plug into your observability, and treat like any other internal service.
Transcript
Justy So Anthropic just quietly dropped a self‑hosted version of Claude Managed Agents, and the pitch is basically: you can run the whole agent runtime behind your own firewall.
Cody Yeah, not just the sandbox this time. They’re saying you host the harness, the database, the workers, the credentials, the session history… the whole thing lives in your cloud or on‑prem.
Justy And you still choose where the model calls go. So Claude could be hit directly, or via Amazon Bedrock, Google’s Agent Platform, Microsoft Foundry, or your own gateway.
Cody Right.
Justy Okay, quick aside, because this feels like the most “one‑year‑in” Exploring Next move ever: remember when we were arguing whether the agent harness would end up inside the cloud platforms or as its own thing?
Cody Mm‑hm.
Justy This is Anthropic kind of trying to do both. It’s their harness, but now you can drop it into the same V P C where you run everything else.
Cody The central claim in the article is pretty clean: some teams want agents that touch real code and production systems, but they can’t use a hosted control plane. Until now they basically had to build their own harness. Self‑hosted Managed Agents is Anthropic saying, “install ours instead, keep data and credentials on your side.”
Justy And the examples they give are all high‑stakes, unattended stuff: fixing broken builds, testing releases, triaging alerts, chewing through a bug backlog. That’s exactly the work where security and audit freak people out.
Cody Yeah. The interesting part to me is the security architecture they lay out. They lean on four principles, and those actually matter more than the self‑hosting label.
Justy Walk through those?
Cody First, they split what they call the brain and the hands. The harness is the brain: it calls Claude, manages the agent loop, keeps session history, holds vault secrets. The sandbox is the hands: that’s where the agent’s code runs. Those are deliberately separate. So the sandbox can die, be recycled, whatever, and you don’t lose history or secrets because they never lived there in the first place.
Justy So the agent’s long‑term memory and its keys are in the harness, and the sandbox is just a throwaway workspace.
Cody Exactly. Second, credentials stay out of the sandbox. They’ve got vaults, and for M C P tools or supported API calls, a proxy outside the sandbox injects the credential. With self‑hosted, you can back that with your own secrets manager instead of Anthropic’s.
Justy That’s the bit I think real teams will care about. If your security review already blessed your internal secrets manager, you get to reuse that instead of convincing them about some new cloud vault.
Cody Third, they let you limit network access per environment. So you can set an allow list of hosts and block everything else. Their example is an agent that fixes builds only being allowed to hit your git host and your package registry. If you host the sandbox, you enforce the same thing with your firewall rules.
Justy So no “oh, the build‑fixing agent randomly posted something to a random webhook” surprises, at least in theory.
Cody In theory. The last one is observability: every session leaves a detailed record of steps, tool calls, and responses, and they export that over OpenTelemetry. So you can pipe it into whatever you already use for logs and traces and look at agent activity next to your other services.
Justy That’s a nice alignment with the whole ‘agents as distributed systems’ thing we talked about a few episodes ago. If you can’t see what it did, you can’t debug or trust it.
Cody Yeah, and OpenTelemetry is the right choice there. You don’t want Yet Another Custom Log Viewer just for agents.
Justy So, on their own argument: do you buy that this actually makes agents safe enough to run unattended on real systems, or is this mostly a compliance story?
Cody The architecture is legit — splitting brain and hands, externalizing secrets, tightening egress, and having a full event log are all the right moves. But it doesn’t magically solve agent behavior. If you prompt an agent to “fix all flaky tests tonight,” it can still call the wrong tool in the right sandbox. You’ve limited blast radius, not eliminated mistakes.
Justy So the security story is “we wrapped the chainsaw in a safer casing,” but you can still cut the wrong tree down.
Cody Yeah. And self‑hosting adds operational work. You’re now running the Managed Agents server, its database, and the workers. You get control, but you also get pager duty when the harness pod crashes in the middle of a ten‑hour run.
Justy From the product side, I see this very clearly aimed at big enterprises who were already almost sold on Managed Agents but blocked by two things: data residency and ‘no external control plane’ rules. If your legal team keeps saying, “we can’t ship a service where a vendor stores session history for our production agent,” this is the path around that.
Cody Yeah, or if you’re on, say, Azure and you want to standardize on Foundry for all outbound model calls, they explicitly support routing from the self‑hosted harness to Foundry instead of directly to Anthropic. That’s a big integration knob.
Justy Who do you think shouldn’t touch this yet?
Cody If you’re a smaller team who just wants an agent to help with docs or light code review, the hosted Managed Agents are probably easier. Spinning up your own agent control plane, wiring up vaults and firewalls and OpenTelemetry, that’s real work. You only take that on when compliance or internal‑network access actually demand it.
Justy Yeah, if your threat model is basically ‘don’t leak customer emails,’ the hosted thing with zero data retention is already pretty compelling. The self‑hosted beta is more like: you’re ready to let agents touch prod, and your infra and security folks insist everything live inside your perimeter.
Cody And even then, there are open questions the article doesn’t answer: pricing is by quote, there are hints elsewhere that some hosted features don’t carry over yet, and you still have to integrate this into your existing auth and review flows. So it’s not plug‑and‑play safety; it’s a more deployable harness.
Justy Still, as a direction, I like it. It treats the harness as infrastructure you can run wherever your real systems live, instead of a black box you have to trust from afar.
Cody Yeah, that part I’m fully on board with. If agents are going to be long‑running services, they should look and feel like the rest of your stack — same V P C, same secrets manager, same observability — not a magical SaaS sidebar.
Justy Alright, Cody, I’m calling it: this is the least flashy, most consequential type of launch. Let’s see who actually signs up for the beta before we argue about it again.