Priyanka Phatak and Gabriel Cemaj introduce a workshop on Claude Managed Agents. Phatak describes an initial architecture that placed the session log, harness and sandbox in one container. This simplified local execution but made debugging, recovery and private-network integration difficult, because the container became state that could not be safely discarded.
The revised design separates an append-only session log, a model-and-tool execution loop and on-demand sandboxes or external tools. Phatak explains how a harness can resume from the event log and how sandbox failure becomes an error the agent can handle. A proxy injects credentials outside the sandbox, reducing direct credential exposure; this architectural measure should not be read as a guarantee against every prompt-injection risk. She reports faster time to first token without presenting an independent benchmark.
The introduction concludes with environment, agent and session APIs. Environments define packages and network policy, agents define prompts and tools, and sessions exchange user messages, custom-tool events and execution status. These interfaces aim to outlast model-specific harness assumptions, which the presenters say must be reevaluated as models change.
Watch on YouTube
