DEEP_DIVE_STORY.LNK

Anthropic introduced a decoupled Managed Agents service that stabilizes agent interfaces while letting harnesses and sandboxes evolve.

Anthropic’s new Managed Agents virtualize an agent into three parts—session, harness, and sandbox—so you can swap internals without breaking the outer contract. The team calls out how harness assumptions go stale as models improve (e.g., context resets needed for Sonnet 4.5 became dead weight on Opus 4.5), and builds around stable interfaces to keep long-horizon runs reliable. Read the engineering deep dive on how they designed for “programs as yet unthought of” and why pets-vs-cattle applies here Anthropic.

At the integration layer, momentum is building around the Model Context Protocol as a standard substrate for agent connectivity. Coverage shows Amazon leaning into MCP for agentic workloads The New Stack, and community posts walk through practical MCP adoption paths DEV.

Evidence continues to favor “skills” and structured workflows over free-roaming agents: a research evaluation framework reports ~20% absolute accuracy gains when relevant skills are activated, but warns activation often drops to ~40% without guardrails Tessl. Fresh skill catalogs are also shipping, including a release that expands installable skills to 1,412 with curated LambdaTest automation bundles GitHub. Ops advice focuses on persistent execution logs, state checks before retries, and a policy layer between agents and external APIs Adopt.ai.