Tech • AI • Robotics • Game
OpenAI's new secret weapon has just been released
8/10TL;DR
OpenAI has launched a public beta Agents API that exposes the managed Codex runtime behind autonomous developer agents, shifting competition from model quality alone to orchestration, tools and execution infrastructure.
KEY POINTS
A managed agent runtime, not a simple model endpoint
The release on September 10, 2026 is centered on the harness behind Codex, not just access to a language model. In practical terms, developers define an agent, its instructions and tools, while OpenAI runs the loop that plans, calls tools, reacts to errors and continues until a task is completed. That marks a move from prompt-response usage to long-running task execution.
Why the harness matters
In an agent stack, the model is only one component alongside the harness, tools and execution environment. The harness handles context management, permissions, sessions, retries, summarization when context fills up, sub-agent orchestration and execution flow. That means a stronger harness can improve an agent even without changing the underlying model, and model benchmarks alone no longer predict which coding agent performs best.
Four core building blocks
The API is organized around agents, environments, sessions and events. A session is durable, keeps history and work state, and can be resumed or redirected while running. Developers therefore interact with a persistent task object rather than rebuilding orchestration logic for every call.
Built for tasks that last minutes or hours
A central capability is persistent sessions that keep working through failures instead of stopping at the first error. An agent can read code, run tests, make changes, detect a failure, fix it and retry before returning a result. OpenAI says these sessions can run for minutes, hours or longer, which makes them closer to background workers than chatbots.
Automatic context compaction
Long sessions accumulate file reads, command outputs and errors until context windows become saturated. The harness automatically compacts prior work into a summary of completed steps, pending work and decisions taken, then continues with a lighter context. That removes one of the most difficult engineering problems teams usually face when building agents themselves.
Parallel sub-agents
With multi-agent mode enabled, a main agent can spawn specialized sub-agents for areas such as backend analysis, security review or testing, then merge their findings. By default, the system allows six simultaneous sub-agents, though that can be adjusted. Their contexts are isolated, but they share the same filesystem, which means concurrent edits to one file still need coordination.
Tool Search and programmatic tool calling
The platform addresses tool overload with Tool Search, which loads only relevant tool definitions on demand instead of injecting potentially hundreds into context. It also enables programmatic tool calling by default: the model writes a small JavaScript program that can call multiple tools in parallel, filter and merge outputs, and send back only useful results. The result is fewer model passes, lower token use and less raw data pushed into context.
Execution can run in three places
The harness is managed by OpenAI, but the execution environment is separate. Developers can use the hosted Sandbox, run a local or self-hosted executor that pulls commands over an outbound WebSocket connection, or rely on third-party providers such as Cloudflare, Vercel, Modal or Daytona. This separation is important for deployment choices and security boundaries.
Sandboxing does not solve capability risk
The hosted Sandbox provides an isolated Linux machine per session with configurable packages, startup commands, files, environment variables and downloadable output artifacts. Network policy can be open, restricted by allowlist, or disabled. But shell isolation does not prevent dangerous actions if the agent has powerful permissions such as write access to GitHub, AWS or Stripe, or if it is manipulated by prompt injection through untrusted external content.
Pricing, limits and strategic stakes
There is no separate fee for the orchestration layer itself, but usage still incurs model, tool and container costs. Reported pricing includes $10 per million input tokens, $1 per million cached input tokens, and $50 per million output tokens for GPT-6 Astra, plus tool and sandbox charges. Beta limits include US-only data residency, no zero data retention even for self-hosted execution, and a required beta header. Strategically, the launch signals that OpenAI is no longer selling only model access, but a managed platform on which outside agents can run.
CONCLUSION
The new Agents API turns Codex from a product into infrastructure, giving developers a managed runtime for durable, tool-using and multi-step agents. Its significance lies less in raw model access than in the attempt to own the operational layer that makes AI agents reliable at scale.
Explain this