Anthropic Shifts from Model Sales to Managed Agents: The Rise of Agent-as-a-Service

Anthropic announced the public test of Claude Managed Agents, a hosted platform where users define tasks and tools and Anthropic runs the agents with all infrastructure included.

"Anthropic wants to sell agents, not only the models. It's a huge market, and it will change the pricing structure away from tokens."

Historically, large‑model companies have sold raw model calls billed per token, a business model that faces shrinking margins as token prices drop. Managed Agents introduces a new model: selling complete agent runtime services, similar to the cloud evolution from virtual machines (EC2) to managed functions (Lambda).

This shift promises higher per‑agent revenue and stronger customer lock‑in, because enterprises that embed business processes in the agent platform face high migration costs.

"Building Managed Agents meant solving an old problem in computing: how to design a system for 'programs as yet unthought of.'"

Anthropic’s solution splits an agent into three independent layers:

Brain layer : the Claude model plus a Harness execution framework that makes decisions without caring where tools run.

Execution layer : sandbox containers, MCP servers, and external APIs where the actual work happens; a single agent can control multiple execution environments simultaneously.

Memory layer : an event‑stream store that records every step, decoupled from the brain and execution layers so that loss of any component does not erase history.

The three layers can be scaled independently: the brain layer is stateless and horizontally scalable, the execution layer is allocated on demand, and the memory layer provides persistent storage for replay.

Performance measurements show a 60% reduction in p50 token‑response latency and over 90% reduction in p95 latency, indicating a substantial user‑experience improvement.

The UI offers a templated creation flow with a "What do you want to build?" prompt and pre‑built agent templates such as Deep research, RAG retrieval, Structured extraction, Intent router, Draft generator, Reflection loop, Task planner, and Summarizer. Users can also start from a blank canvas, upload documents as context, add sub‑agents, or let Claude design the agent structure via conversation.

Security is a core principle: all credentials are stored outside the sandbox in a secure vault. For example, Git tokens are injected into the sandbox’s git remote configuration at initialization, and MCP OAuth tokens are passed through a dedicated secure channel, preventing credential leakage even if code execution is compromised.

Anthropic has already piloted Managed Agents with enterprise customers such as Notion (parallel task execution for code delivery and content generation), Rakuten (multi‑department agent deployment with Slack and Teams integration), Asana (an "AI teammate" that collaborates within project workflows), and Sentry (a debugging‑and‑patching agent that automates bug detection to PR submission). These deployments achieve dramatically shorter rollout times—days to weeks instead of months.

Strategically, Anthropic is transitioning from a pure model provider to an agent platform company, a move echoed by OpenAI’s Assistants API and Google’s Vertex AI Agent Builder, but Anthropic’s approach fully hosts the agent runtime rather than adding a thin layer on existing platforms.

For developers, this reduces infrastructure complexity, allowing focus on business logic. For startups building agent‑orchestration layers, the market becomes tougher as model vendors offer end‑to‑end solutions.

Managed Agents remain in public testing; multi‑agent coordination and self‑evaluation features require a research preview request, and the product is not yet fully mature, but the direction appears irreversible.