Agent platforms are fragmenting

Artificial intelligence “agents” are turning into infrastructure products, and the split is happening in the control layer, not just the model. Anthropic and Amazon Web Services now pitch managed agent systems with different defaults for tools, memory, and governance. (platform.claude.com) (docs.aws.amazon.com) An agent is software that can plan steps, call tools, and keep working across a session instead of answering one prompt at a time. Anthropic’s Claude Managed Agents says it handles the agent loop, tool execution, runtime, web browsing, code execution, and file work inside a managed environment. (platform.claude.com 1) (platform.claude.com 2) Amazon Web Services is selling a broader platform layer around the same idea. Amazon Bedrock AgentCore says it can deploy and operate agents “using any framework and model,” with separate services for runtime, memory, identity, gateway access to tools and data, and observability. (docs.aws.amazon.com) (aws.amazon.com) (docs.aws.amazon.com) That leaves developers choosing between a hosted harness and a composable platform. Anthropic’s setup starts with a reusable agent definition that bundles a model, system prompt, tools, Model Context Protocol servers, and skills, while Amazon Bedrock also offers inline agents configured at runtime with action groups, guardrails, and knowledge bases. (platform.claude.com) (docs.aws.amazon.com) The technical divide is showing up in standards too. Anthropic introduced the Model Context Protocol on November 25, 2024 as an open standard for connecting assistants to data and tools, and in December 2025 said it was donating the protocol to the Linux Foundation’s Agentic AI Foundation with support from Amazon Web Services, Google, Microsoft, and others. (anthropic.com 1) (anthropic.com 2) Model Context Protocol works like a universal plug shape for agent tools, but it does not tell companies how to run the agent itself. The protocol documentation says it defines client-server context exchange and leaves application behavior and model orchestration outside the spec. (modelcontextprotocol.io 1) (modelcontextprotocol.io 2) Amazon is putting more weight on enterprise controls around those connections. Its documentation says Bedrock Guardrails can be applied across text and image models and can also be used with Bedrock Agents and Bedrock Knowledge Bases. (docs.aws.amazon.com) (docs.aws.amazon.com) Anthropic is putting more weight on the hosted runtime itself. In an engineering post published 4 days ago, the company said managed agents are built around stable interfaces while the underlying harness changes as models improve, a design meant for “long-horizon” work that runs over longer sessions. (anthropic.com) The result is not one market for “agents” but several adjacent ones: model vendors selling managed autonomy, cloud providers selling governance-heavy platforms, and standards groups trying to keep tool connections portable. The fragmentation is happening above the model, where companies decide who controls the loop, the tools, and the rules. (platform.claude.com) (aws.amazon.com) (modelcontextprotocol.io)