The Agent API gives developers the execution infrastructure used by Codex: long sessions, sub-agents, hosted sandboxes, no charge beyond consumed tokens.
Announced on September 10, the OpenAI Agent API responds to the most common complaint among those who deploy agents: most of the work is not in the model, it is in the plumbing around. Resuming an interrupted session, compacting a context that overflows, isolating code execution, coordinating several agents between them — it is this layer that OpenAI now offers in service, opening access to the harness that runs Codex internally.
The announced perimeter covers long-term sessions, the call of tools via MCP or home tools, the coordination of sub-agents, context compaction on extended tasks, and execution in a sandbox hosted by OpenAI or in a computing environment chosen by the developer. Orchestration, session management and context management are on the supplier's side.
The public beta is open to all API developers, at no extra cost: billing only covers tokens and tools consumed by agents. OpenAI highlights the results of its early users — latency divided by four, cost per task down 60%, 86% less unsuccessful responses — figures that come from the publisher and depend largely on what a home orchestration is compared to.
Read the full article on the original source
Go to the source →