If you're comparing Claude Managed Agents with Wallfacer, start with what Anthropic got right, because a lot of it is very good.
Managed Agents does the hard infrastructure well
Anthropic shipped Claude Managed Agents in April, built on the idea of decoupling the brain from the hands. You get a hosted agent and a sandbox without running any servers.
Secrets are swapped in on the way out of the sandbox, so the agent never sees them. Sessions can have a hard spend limit, and agents are versioned, so you can pin the one you tested.
You can limit a sandbox's outbound traffic to a list of hosts, or run the sandboxes on your own machines. Pricing is published: token prices plus 8 cents per session-hour.
Anthropic also ships working recipes, like a Slack bridge and an incident responder that takes an alert all the way to a pull request.
A lot of it is the same in Wallfacer
Wallfacer is built on the same split. The model's loop runs in one machine and your code in another, and neither product asks you to run servers.
Both connect to MCP servers and keep secret values out of the transcript. Both give agents memory that carries across sessions, run work on a schedule, and keep a full log of every session.
Most of the setup is the same work too. Either way, someone gives the agent access to the repository, a toolchain that builds it, and the secrets it needs.
The differences show up on a team
Managed Agents leaves the agent's place on a team to the application you build: who it is, what it can reach, how work gets to it, and who signs off. If you're building agents into your own product, that's the right split.
Those are also where teams run into trouble. One developer on Hacker News said reviewing Claude's merge requests took five times as long. Microsoft's .NET team saw their agent's success rate go from 38.1% to 69% once it could build and test the project.
Wallfacer builds that team layer in.
Wallfacer agents show up as themselves
On Managed Agents, the agent acts on GitHub as whoever's token you give it, and Anthropic's GitHub guide asks for a personal access token. Unless you set up a separate account, its pull requests carry your name. The Claude Code Routines docs put it plainly: anything a routine does "appears as you."
Every Wallfacer agent has its own GitHub account, its own Slack app, and its own email address. When it opens a pull request, the pull request is from the agent. When it replies in Slack, the agent's name is on the message.
Your reviewers know what they're looking at, and the record keeps you and the agent apart. If you offboard an agent, its history stays attributed to it.
Anthropic's Slack product, Claude Tag, takes a different approach: one GitHub App for the whole organization. A commenter in its launch thread pointed out that the audit log can't tell which channel's Claude did something.
Wallfacer agents only carry their own keys
Each Wallfacer agent gets its own GitHub token, model key, Google account, and the logins you've granted it, and nothing else. A support agent with a Zendesk key and no GitHub account has no way into your repos. A reviewer with read-only access can't push.
On Managed Agents, credentials come along with whichever session your code starts, so keeping them straight is your code's job. For Anthropic's own cloud sessions, better secrets handling is the most-reacted request we found, with 212.
With Wallfacer, work starts where it already happens
Nobody should have to remember to go type a prompt. Assign a Wallfacer agent a GitHub issue, mention it in Slack, or email it, and it gets to work. Anything that can send a webhook can start a playbook too.
With Managed Agents, a session starts when your code calls the API, so each of those paths is something you build and host. Anthropic publishes a Slack bridge to start from. Over on Claude Code, assigning Linear issues to Claude is a request with 148 reactions.
Wallfacer has an agent review the work before a person does
Review is where teams feel agents the most, so we made it part of the process instead of something someone has to remember.
When an engineering agent opens a pull request, a reviewer agent goes over it first, and it can run a different model from the author. Then the task waits for a person to approve it on GitHub. If they ask for changes, the work goes back to the agent with their feedback.
In August, on one of our own projects, the tests passed and our reviewer agent still caught two real bugs: duplicate person records, and approvers contacted through the wrong channel. The fix came back, got approved, and then a person merged it.
Managed Agents can pause a single tool call for approval. A review process like this one is something you'd assemble yourself.
Wallfacer agents learn how your team works
Instruction files help, but people say they don't always stick. In Wallfacer, the way your team works lives in a handbook you own, and each agent has a role page that reads like a job description. Every task the agent runs starts with both.
When a review turns up a gap, the fix is usually a sentence on a handbook page, and the next task picks it up.
Wallfacer agents can run the app they're changing
Agents do much better when they can build and test their own changes. A Wallfacer agent's computer has your project set up and your services running, boots from a snapshot, and lets you preview the app. On macOS it has Xcode and the iOS Simulator, so a mobile change can be built and looked at.
Managed Agents gives each session a fresh Linux container with up to 8 GB of memory, and you start what you need inside it each time.
Where Wallfacer still has work to do
A Wallfacer agent's credentials live in its shell, where Managed Agents keeps secrets out of the sandbox. Our computers have open outbound access, and Managed Agents can limit it to a list of hosts.
Running Wallfacer on your own hardware is an enterprise deployment rather than self-serve, and our prices aren't public yet.
A person still reviews the work, and our triggers are judged by a model, so they're flexible but not exact. We have completed our SOC 2 Type II audit.
So which one?
If you're building agents into your own product for your own customers, Managed Agents gives you great building blocks. If you want agents your team can hand work to like a colleague, under their own names, with their own access and your process around them, that's what we built Wallfacer for.
The full side-by-side is in Wallfacer and Claude Managed Agents.