Evolution of Agentic Surfaces

Long-running agents need to survive tool failures and resume without losing their work. Anthropic Applied AI engineers Gagan Bhat and Isabella Kai He explain how Claude Managed Agents separates model reasoning, tool execution, and session history. They also argue that a harness—the software around a model that manages its tools and context—must change when its model-specific workarounds stop helping.

Key Points Covered

• Production agents require more than an agent loop: Bhat lists hosting, session state, file access, execution isolation, credentials, and observability as infrastructure that teams otherwise build around model calls and tools 01:54–05:45(opens in a new tab).
• Managed infrastructure changes the ownership boundary: Anthropic's proposed split leaves the product, task, context, and domain knowledge with the developer while the service owns the model loop, sandbox, session management, credentials, observability, and hosting 05:45–06:43(opens in a new tab).
• Harness assumptions expire as models change: Kai He describes context resets added for Sonnet 4.5's early stopping near its context limit. When Opus 4.5 no longer showed that behavior, the workaround added latency and cache problems instead 07:31–09:19(opens in a new tab).
• Small replaceable primitives reduce migration drag: Kai He argues for independently changeable harness components so model-specific behavior can evolve without replacing the overall architecture 09:19–10:13(opens in a new tab).
• Long-running agents need context, isolation, recovery, and parallelism: The speakers frame secure sandboxes, context engineering, reliable hours-long execution, and concurrent work as baseline harness requirements rather than application features 10:13–11:07(opens in a new tab).
• Separate the brain from the hands: Moving the model loop out of the tool-execution container lets reasoning begin before a sandbox is ready, limits shared failure modes, and allows a failed sandbox to be replaced without losing the agent's state 11:07–12:55(opens in a new tab).
• Make the session a durable resource: Their agent, environment, and session primitives store interactions outside the active context window. A restarted loop can resume from the log, and discarded context can be read back into a later window 12:55–16:13(opens in a new tab).
• Keep credentials outside model-visible context: The product's vault design decrypts credentials only during tool execution. Self-hosted sandboxes and MCP tunnels aim to keep execution and private services within customer-controlled boundaries 21:26–22:18(opens in a new tab), 25:07–26:45(opens in a new tab).
• Decoupling can improve startup latency: Anthropic reports 60% faster P50 and more than 90% faster P95 time to first token by starting model reasoning and sandbox setup in parallel. The talk does not provide the workload distribution or experimental details needed to transfer those figures 22:18–24:17(opens in a new tab).
• Logs can support observability, memory, and evaluation: The speakers propose using session traces both for inspection and for periodic memory updates, while a separate outcome grader checks a user-defined rubric and triggers retries. These mechanisms still depend on the quality of the extracted memory, rubric, and grader 24:17–25:07(opens in a new tab), 27:34–30:22(opens in a new tab).

Full video: https://www.youtube.com/watch?v=K0X9QDRkIdg(opens in a new tab)