Battle over the AI 'harness'

An AI agent still needs a control layer to call tools, manage memory, and run code — and big labs are now selling that layer in different ways. (thenewstack.io) Anthropic put that layer into public beta on April 8 as Claude Managed Agents, a hosted service that runs long tasks in managed infrastructure. Anthropic’s docs say it provides the “agent loop, tool execution, and runtime” so developers do not have to build them themselves. (platform.claude.com) Anthropic bills that service at $0.08 per session-hour on top of model token charges, with separate pricing for features like web search. The New Stack reported that pricing as part of a broader split that became visible after OpenAI updated its own tooling seven days later. (platform.claude.com, thenewstack.io) OpenAI took the opposite route. Its Agents SDK docs describe a higher-level runtime with an “agent loop” that manages tool invocation, handoffs, sessions, guardrails, and sandbox agents, and the code is published as an open-source project on GitHub. (openai.github.io, github.com) That turns a technical plumbing choice into a business fight. Anthropic is charging for orchestration on its own infrastructure, while OpenAI is treating orchestration as software that can spread freely and pull usage back to its models and tools. (thenewstack.io, platform.claude.com) The harness is the part users rarely see but enterprises care about first. Anthropic’s engineering team describes it as the loop that calls the model and routes tool calls, while OpenAI says its SDK is for teams that want the runtime to manage turns, tool execution, sessions, and artifacts across multiple steps. (anthropic.com, openai.github.io) Other vendors are building the same stack. Google’s Vertex AI Agent Engine offers sessions, memory across sessions, and managed code execution, and Microsoft’s Foundry Agent Service offers managed long-term memory for continuity across sessions and workflows. (docs.cloud.google.com, discuss.google.dev, learn.microsoft.com) The term itself has spread fast in 2026. Martin Fowler’s site defined “harness” this month as everything in an AI agent except the model itself, and OpenAI tied the idea to a February post about building a million-line internal system with agent-driven code generation. (martinfowler.com, openai.com) Anthropic is also arguing that the layer should stay flexible because model behavior changes. Its engineering blog says harness assumptions “go stale as models improve,” which is one reason the company built Managed Agents around stable interfaces like sessions, harnesses, and sandboxes. (anthropic.com) The next test is whether companies pay a separate meter for that control layer or expect it to come bundled with model access. The code may be invisible to end users, but the pricing now sits in plain view. (thenewstack.io)