Every major agent platform is now selling the same thing: the loop that sits between a model and a working application. OpenAI’s Agents API, launched last week, gives developers direct access to the managed Codex harness that used to be an internal implementation detail. Google Cloud shipped an agent plugin around the same window. Analyst Josh Rosen argues on X that both moves are symptoms of a bigger shift: the harness, not the model, is becoming the product.
The pitch from the model labs is straightforward. OpenAI says new model capabilities frequently require harness changes to use well, so it intends to ship the two together. Anthropic makes the same case for Claude Managed Agents. The implied argument is that controlling both halves produces a better outcome than either one alone: sharper harnesses because the lab knows the model’s quirks, and arguably sharper models because the lab knows what the harness will demand of them.
AWS and Microsoft are chasing the same layer from a different angle. Neither company owns a frontier model the way OpenAI and Anthropic do, so AgentCore Harness and Foundry Agent Service are built to work underneath whichever model a customer picks, Bedrock, OpenAI, or Gemini included. Vercel’s AI SDK approaches the problem from a third direction, wrapping infrastructure around agent harnesses rather than issuing one itself. That split, model-tied versus model-agnostic, is the first fork any team building an agent now has to navigate.
Six capabilities are getting absorbed into these managed layers, and each one used to be something a team wrote itself. A harness can now be declared in a single API call rather than assembled as custom orchestration code, the way Kubernetes moved infrastructure from scripts into configuration. Agent versions get immutable releases and named endpoints, with AWS running A/B tests across versions on live production traffic. The model attached to a session is no longer fixed at build time: AWS lets a team swap models mid-conversation, and Microsoft’s router can send a simple turn to a cheap model and a hard one to a stronger model, all without touching the application code. Tools and skills are moving into shared infrastructure too, reachable through a gateway or a toolbox rather than wired into each agent individually, so a company builds GitHub or Salesforce access once instead of fifty times. Microsoft’s Agent Optimizer, and a similar system AWS is assembling inside AgentCore, close the loop further by generating alternative agent configurations from production traces and testing them before a human promotes one.
That is a lot of decision-making handed to a vendor, and it is worth being direct about who is making each claim. OpenAI and Anthropic are describing their own infrastructure in their own launch materials; there is no independent benchmark here comparing a self-built harness against a managed one on cost, latency, or reliability. The argument that co-developed models and harnesses beat a general-purpose harness is a reasonable hypothesis, not a demonstrated result.
The practical question for any team shipping an agent this quarter is narrower than “build or buy.” It is which pieces of the loop are truly generic, worth outsourcing to whoever runs it best, and which pieces encode something specific to the product: a workflow, a proprietary data source, a constraint a general harness will never model correctly. Orchestration and versioning look close to commodity already. The tool integrations and skills tied to a company’s own systems look far less so.
Teams choosing between OpenAI’s harness, Anthropic’s, or AWS and Microsoft’s model-agnostic alternatives in the next two quarters should treat the model-lock-in question as the deciding variable, not the feature list. A managed harness tied to one lab’s models saves engineering time until a better model ships elsewhere, at which point the switching cost is the whole harness, not just an API call.
Based on analysis by Josh Rosen, published on X on September 11, 2026.
