The AI agent harness — the control layer surrounding a model that handles orchestration, memory, sandboxed execution, and observability — has become the new battleground for frontier labs. Anthropic launched Managed Agents in public beta at $0.08 per session hour. OpenAI countered by shipping an open-source Agents SDK with no additional runtime fee, charging only for model tokens and tool calls. Google meters harness components separately via Vertex AI Agent Engine, while Microsoft uses consumption-based billing in Foundry Agent Service. AWS is co-developing a Stateful Runtime Environment with OpenAI for Bedrock. The strategic split: Anthropic, Google, and Microsoft charge for the runtime in various forms; OpenAI gives it away to drive model consumption. This creates two compression forces on startups — free open-source from OpenAI squeezes horizontal orchestration frameworks like LangChain and CrewAI, while paid managed services from the hyperscalers squeeze paid harness platforms. The defensible startup position is differentiation into governance, compliance, vertical depth, or multi-model control rather than horizontal orchestration.
Table of contents
What a harness is, and why it became a marketWhat Anthropic shipped, and what it costsWhat OpenAI shipped, and what it costsThe labs agree on owning the layer. They disagree on how to bill it.The middleware arc, with a split running through itWhat this means for startups that filled the gapWhat this means for teams that built their ownWhat’s nextSort: