Can a Managed Agents Worker Pass Your Canary?
Keep the Claude Managed Agents definition fixed, run a deterministic canary through a new self-hosted environment, and compare stage timing and cost per passing run with the cloud environment.
Exercise the worker with a result you can check
Keep the agent definition fixed across the cloud and self-hosted trials. Start new sessions against the appropriate environment IDs so the environment is the deliberate change. On a synthetic fixture, have the tool count the three lines in `red\nblue\nred` and return the exact text `lines=3`. Capture session creation, first tool start, tool completion, and result arrival using one timing origin per run. For the self-hosted trial, compare the worker’s completion record with the result the caller observes. A tool that finishes in 20 milliseconds but whose result arrives 900 milliseconds later points to a different investigation from a tool that never starts. Use the documented environment-key flow for the applicable platform, checking the AWS variation before provisioning credentials.
Choosing an agent environment - Claude Platform DocsCompare what the two runtimes actually keep local
Use the placement of the agent loop as the first comparison. With Claude Managed Agents, your worker executes bash and file tools locally while Anthropic determines which tool to call. Cursor Self-Hosted Machines makes a comparable split for coding jobs, with the agent loop remaining in Cursor’s cloud. A local checkout requirement can fit either design. A policy that requires inference to stay within your boundary would rule out both documented arrangements. For cost, compare the same canary in Claude’s two environments and allocate worker idle time across the self-hosted trial. An illustrative $3.60 spent across 12 attempts with nine passing on time is $0.40 per pass. A cloud trial at $3.00 with ten on-time passes costs $0.30 per pass, before assigning value to any placement requirement.
Choose where Cloud Agents run | Cursor DocsTest the requirement that changes the environment
Claude Managed Agents offers an Anthropic cloud environment with a fresh Linux container per session and a self-hosted sandbox where your worker executes bash and file tools. The same agent definition can target either environment. Claude still chooses tool calls on Anthropic's side and receives the worker's results, so review what each tool may return before granting it access to sensitive inputs.
In an illustrative pilot, a build requires a private compiler and the only approved result is a pass or fail status. Run the same build against a synthetic fixture in the cloud environment and on your worker. Check where the command executed and inspect the returned tool output. Choose self-hosting if it meets a requirement the cloud environment cannot satisfy and the tool output stays within policy.
Choosing an agent environment - Claude Platform Docs