OpenAI builds multi-agent orchestration directly into its API
Programmatic Tool Calling lets the model write JavaScript to coordinate its own tools. Multi-agent spins up concurrent subagents inside one request. The framework layer just got absorbed.
- Programmatic Tool Calling is generally available in OpenAI's Responses API: the model writes JavaScript that runs in a hosted, isolated V8 sandbox to chain tools with loops, conditions, and parallel calls.
- Multi-agent is in beta: a root agent spins up concurrent subagents and synthesizes their work inside a single API request.
- Orchestration is moving from external frameworks into the platform layer, which reprices the scaffolding market.
Buried in OpenAI's busy July 9 was the part that matters most to this beat: two orchestration primitives shipped in the Responses API, announced by the OpenAI developer team. Programmatic Tool Calling is generally available, and Multi-agent is in beta. Together they move a layer of the agent stack that used to require an external framework inside the API itself.
The model writes its own glue code
Programmatic Tool Calling lets the model write JavaScript that runs in a hosted V8 sandbox and coordinates its tools with loops, conditions, and parallel calls, instead of issuing one tool call per turn and waiting. The sandbox is deliberately austere: a fresh, isolated runtime per execution, top-level await, no Node.js, no package installs, no direct network access, and no state that persists between runs.
Multi-agent, the beta half, lets a root agent delegate to subagents that run concurrently and get synthesized into one answer within a single API request. OpenAI positions it for codebase exploration, documentation, and implementation work, the same fan-out pattern teams have been hand-building all year.
What happens to the framework market
Every team running LangGraph or CrewAI-style scaffolding just got a build-versus-buy question they did not have last week, and the vendors selling that scaffolding got a new competitor: the API bill itself. It is the same squeeze Microsoft applied when it merged AutoGen and Semantic Kernel into one SDK, and the counter-move is the one LangChain made by going down-stack with NVIDIA on cost. Orchestration keeps consolidating toward whoever already bills you.
Worth remembering the meter runs on all of it: parallel subagents are a spend multiplier on top of the pricing ladder OpenAI shipped the same day. The orchestration is free. The tokens are not.