The AI model gets the credit. What is doing the rest of the work?

A model response is one part of an agent. Something also has to decide which tool to call, record the result and ask for the next step. Anthropic’s Managed Agents architecture describes three separate components: a session holding the event history, a harness that calls the model and routes actions, and a sandbox where code runs or files change.

Keeping them separate solves an ordinary infrastructure problem. When everything lives inside one container, losing that container can mean losing the session as well. A service with stable interfaces can replace an execution environment without treating the entire job as disposable. It can also connect tools or customer environments without making the model loop depend on one physical machine. This does not make the model’s decisions correct; it gives a longer task a more durable record and a clearer place to recover when infrastructure fails.

That separation also makes failures easier to diagnose. A stalled run can come from an expired credential, a lost process, an incomplete tool result or a model decision. Keeping event logs and checkpoints lets an operator resume work and identify which layer needs repair, instead of treating every failure as evidence that the model itself is incapable.

A separate March experiment from Anthropic examined the organisation of application-building work. Its planner, generator and evaluator roles divided a large request into manageable parts, produced changes and inspected the result. Structured handoffs carried state between sessions. The evaluator used a live browser rather than relying solely on the generator’s description of its own output.

That distinction matters because self-assessment can be lenient, especially for subjective work such as interface design. The experiment used concrete criteria and feedback loops to make improvement assessable. It also distinguished a context reset with a handoff from an in-place summary that lets the same session continue. Neither arrangement is a universal recipe: orchestration adds overhead, and model behaviour changes over time. When evaluating an agent product, inspect its recovery, state and verification mechanisms as well as the underlying model. They determine whether a promising response can become a reviewable piece of completed work.