Two Anthropic Applied AI presenters trace a progression from text-only model APIs to agent SDKs and managed runtimes. They identify the production concerns that appear once agents run for many users, including hosting, session management, credentials, code execution, isolation and observability.
The presenters argue that an agent harnessAn AI agent harness is the software framework that packages a model with tools, instructions, context management, execution controls, and user interaction. must change as model capabilities change. They describe one case where a context-reset workaround became overhead after a later model improved, then explain their design choice to separate the agent loop from tool execution and persist session events so work can resume after failuresDurable workflow execution preserves progress so a multi-step process can resume after interruptions using recorded completion state..
A site reliability incident demo illustrates agent definitions, restricted execution environmentsAn AI agent sandbox is an isolated execution environment that limits which files, processes, networks, credentials, and external systems an agent can access., uploaded evidence and durable sessions. The presenters also discuss credential vaults, parallel container setup, context recovery from session logsContext engineering designs how relevant instructions, evidence, state and tool information are assembled for a model's task., private tool connections, periodic memory updates and a separate grader that checks task outcomesAgent completion verification checks evidence that an agent achieved the requested outcome rather than merely producing intermediate work. against a user-defined rubric. Their performance and product claims are presented as their own reported experience.
Watch on YouTube




