Managed Agents vs Self-Hosted Agent Harnesses

What the vendor actually manages

Anthropic's engineering account of its service (April 2026) describes virtualising three components: a session, "the append-only log of everything that happened"; a harness, the loop that calls the model and routes tool calls; and a sandbox, where code runs and files are edited. Keeping them separate means a failed container no longer loses the session, and inference does not wait for a container to start. Anthropic reports that p50 time-to-first-token fell by roughly 60% and p95 by over 90% after the change. Those are vendor-reported figures for its own system.

OpenAI's Agents API documentation uses nearly the same vocabulary of agent, environment, session and events. It says OpenAI "manages sessions, orchestration, context compaction, and recovery" while the application supplies tools and chooses an execution environment: OpenAI-hosted, self-hosted, or none. Google's Agent Runtime is a different kind of product. It is a managed, auto-scaling environment for agents built with ADK, LangChain, LangGraph, LlamaIndex, AG2 or custom code, alongside managed Sessions, Memory Bank, evaluation and code-execution services. The developer still writes and versions the agent loop.

What a self-hosted harness gives back

Running the harness yourself returns control of the parts that most affect agent quality. The Claude Agent SDK exposes the tools, agent loop and context management that power Claude Code as a Python and TypeScript library, with hooks at lifecycle points, a permission system, subagents, sessions that can be resumed or forked, and MCP connections. Open-source frameworks go further by making the loop itself editable and, in several cases, model-independent.

That control matters because the harness, not only the model, shapes results, which is the premise of harness engineering. A managed service makes those decisions for its customers and may revise them between releases, as Anthropic's beta notice says explicitly. A team that needs a frozen, reproducible agent, or a behaviour the vendor does not offer, has to own the harness.

Security and data boundaries

Managed services have a structural security advantage in one respect: credentials can be kept out of the sandbox. Anthropic describes cloning repositories with an access token during sandbox initialisation, and holding OAuth tokens for MCP tools in a vault reached through a proxy, so that code the model writes never sees them. A self-hosted team can build the same pattern but has to build it.

The disadvantage is that state lives with the vendor. Anthropic states that because sessions store conversation history, sandbox state and outputs server-side, Managed Agents is not currently eligible for Zero Data Retention or HIPAA Business Associate Agreement coverage. OpenAI states that its Agents API supports data residency only in the United States and does not support ZDR. Self-hosted sandboxes move tool execution, files and network egress into the customer's environment, but Anthropic's documentation is clear that tool inputs and outputs still flow to its control plane so the model can see results. For workloads where that is unacceptable, only a fully self-hosted harness calling a model API under suitable terms will do.

Cost and features

Managed pricing is simple to state. Claude Managed Agents bills model tokens at standard rates plus $0.08 per session-hour, metered to the millisecond only while a session is running; idle time waiting for a message or a tool confirmation is not charged, and the Batch API discount does not apply. OpenAI bills model usage, tools and hosted containers at their standard rates. In both cases tokens dominate: Anthropic's own worked example of a one-hour session totals $0.705, of which $0.08 is runtime. The saving from self-hosting is therefore mostly not in the runtime fee. It lies in control over caching, model routing and context size, set against engineering and on-call time.

Managed platforms are also where vendors ship features first. Anthropic's May 2026 update added multi-agent orchestration, in which a lead agent delegates to specialists with their own model, prompt and tools, and outcomes, in which a separate grader scores output against a developer-written rubric in its own context window (both public beta), plus a scheduled memory-curation process called dreaming (research preview). Anthropic reports that outcomes improved task success by up to 10 points over a standard prompting loop; this is a vendor-reported result. Each of these can be reproduced in a self-hosted harness, at the cost of building and evaluating it.