OpenAI and Anthropic Build Models and Harnesses Together… Managed Agent APIs Spread

Frontier labs such as OpenAI and Anthropic are moving agent orchestration from application code to managed APIs. OpenAI Agents API and Anthropic Claude Managed Agents are designed to allow developers to create and run agents without directly implementing complex loops.

Declarative Management of Agents and Release Engineering

Agents are created with a single API call specifying task, model, tools, environment, etc. This is similar to defining orchestration through configuration, like Kubernetes Helm charts. AWS AgentCore and Microsoft Foundry also support versioning, A/B testing, immutable endpoints, etc., managing the release lifecycle of agents.

Dynamic Routing of Model Selection and Externalization of Resources

Previously, models were fixed in agent definitions, but now there is a trend toward dynamically switching models within a session based on task complexity or cost. Additionally, tools and skills are managed as shared infrastructure separate from agents, preventing duplicate implementations and reducing context overload.

Integrated Optimization of Models and Harnesses

The core strategy of frontier labs is to build models and harnesses together. OpenAI maintains and improves the Codex harness alongside the model, internally optimizing context compaction and tool usage logic. Developers use only stable interfaces, while providers can freely change internal implementations. However, if specific behaviors that managed harnesses cannot express are required, developers may need to own the agent loop directly.