Claude Managed Agents: agents, environments, sessions

Try 3 questions on this

Pick an answer to see why each option is right or wrong.

1. Your platform team wants Anthropic to run the agent loop and to host the workspace where bash commands, file edits and code execution take place, so that you write no loop code and operate no sandbox infrastructure. Which approach should you take?

   Right. In the game that’s Renown for your Agent Architecture and Orchestration building.

   Not that one: A is right. In the game your Agent Architecture and Orchestration building would lose a little Renown, and this question would come back later.

   The answerA.Claude Managed Agents with a session per run.Managed Agents is the only one of the four approaches that supplies both the harness and the deployment. Anthropic's orchestration layer runs the agent loop, and each session provisions a container that serves as the agent's workspace for bash, file operations and code execution.Your answerB.The SDK Tool Runner with locally registered client tools.The Tool Runner supplies a harness only. It automates the loop over tools the developer registers, but those tools execute in the developer's own process on the developer's own infrastructure — there is no hosted workspace.Your answerC.A manual Messages API loop on the team's own servers.A manual loop supplies neither harness nor deployment. The team would write the while stop_reason == "tool_use" cycle itself and still run every tool on its own servers, which is the maximum amount of code and operations.Your answerD.The Claude Agent SDK deployed to the team's cluster.The Claude Agent SDK is Claude Code as a library. It ships built-in file, bash and search tools and the full loop, but it is still hosted by whoever runs the process, so the team would continue to operate the infrastructure.

   Show the full explanation and doc quotes

   The other three leave hosting to the developer.

   Written from platform.claude.com/docs/en/managed-agents/overview

2. You read that several agent harness products supply the same loop, and must explain to your team what each one does not provide. Your architects want the route that runs inside the process you already operate and ships no tools of its own. Which approach should you take?

   Right. In the game that’s Renown for your Agent Architecture and Orchestration building.

   Not that one: B is right. In the game your Agent Architecture and Orchestration building would lose a little Renown, and this question would come back later.

   Your answerA.Adopt a Claude Managed Agent, whose harness can supply the loop inside a runtime you operate and ships no tools beside the ones you wrote.Managed Agents is the option that supplies deployment as well as a harness: the agent is created once and versioned, and every session runs in a container Anthropic provisions. That is the right choice when you want the runtime handed to you, and the wrong one when you already have a service.The answerB.Adopt the SDK tool runner, whose harness can supply the loop inside the process you operate and ships no built-in tools beside the ones your caller defines.The SDK Tool Runner is a thin helper inside the regular Anthropic SDK, reached at client.beta.messages.tool_runner, that drives the request, execute and loop cycle over tools the developer registers.Your answerC.Adopt the Claude Agent SDK, whose harness can supply the loop inside the process you operate and ships no tools beside the ones your caller already defines.The Claude Agent SDK is a separate product, not a mode of the Anthropic SDK. It brings its own built-in tools, which is valuable when you want them and is overhead when the functions you need are already written.Your answerD.Adopt a dedicated agent endpoint, whose harness can supply the loop inside a runtime you operate and ships no tools beside the ones you wrote.Every one of these harnesses is built on the Messages API. The tool runner issues ordinary POST /v1/messages calls on your behalf, so nothing in this choice changes which endpoint your service talks to.

   Show the full explanation and doc quotes

   It ships no built-in tools, has no filesystem access and no sandbox, and the developer still hosts the process — which is why it adds nothing to a service that already implements the functions and already runs somewhere.

   Written from platform.claude.com/docs/en/agents-and-tools/tool-use/tool-runner

3. Your Managed Agents session builds a valuation workbook, and about a third of runs come back missing a stated discount-rate assumption. Your service downloads each workbook, checks it against a twelve-point quality list by hand, and sends a follow-up user.message naming what is absent, which adds roughly twenty minutes to every run. How should you restructure the review?

   Right. In the game that’s Renown for your Agent Architecture and Orchestration building.

   Not that one: C is right. In the game your Agent Architecture and Orchestration building would lose a little Renown, and this question would come back later.

   Your answerA.Add the twelve criteria to the agent's system prompt so it reviews its own draft before finishing the turn.A system prompt shapes how the agent works, and self-checking does catch some omissions, but the check shares the context that made the omission and there is no mechanism that re-runs the turn when a criterion is missed. The twelve-point list would still be applied by the party least likely to notice its own gap.Your answerB.Upload the criteria as a file and mount it in the sandbox so the agent reads them while it works.Mounting a file gives the agent the criteria to read while it works, which is how reference material is supplied to a sandbox. Reading a standard is not being measured against one: nothing inspects the produced workbook, so a run that quietly skips a criterion still ends and still reaches your reviewer.The answerC.Send a user.define_outcome event whose rubric names the required criteria, so a grader scores each draft and feeds its verdict back.A user.define_outcome event carries a description of the deliverable and a rubric of gradeable criteria. On receipt the harness provisions a grader that evaluates the artifact against that rubric in a separate context window, so the agent's own implementation choices cannot sway the verdict.Your answerD.Score each downloaded workbook with a separate judge call once the session goes idle, then start a fresh session whenever the check fails.Judging a finished artifact with a separate model call is a sound evaluation pattern; it is the wrong shape here because it sits outside the session. The verdict never reaches the agent, and a replacement session begins with an empty filesystem and none of the intermediate work, so each failure costs a full re-run rather than a revision.

   Show the full explanation and doc quotes

   The grader returns an explanation of which criteria passed or failed, that feedback goes back to the agent, and the cycle repeats until the rubric is satisfied or the iteration cap is spent. The hand review becomes an in-session loop the platform runs.

   Written from platform.claude.com/docs/en/managed-agents/define-outcomes