OpenAI's Agents API exposes the managed Codex harness with sandbox execution, MCP, artifacts, and multi-agent delegation.
OpenAI opened the Agents API in public beta on September 10, 2026, bringing the harness and infrastructure that power Codex to developers through a single flexible API. For software developers who have been stitching together models, tools, and sandboxes manually, the release is a structural shift — with security caveats fresh from September's sandbox escape disclosures.
What the Agents API provides
OpenAI hosts and maintains the harness. Developers choose the agent's compute environment: an OpenAI-managed sandbox, their own infrastructure, or a sandbox partner. The API handles sessions, orchestration, context compaction, and recovery — you supply tools and workflows.
Built-in capabilities include:
- Sandbox execution for code
- File editing within configured scopes
- MCP connections to external tools
- Artifact generation (documents, images, outputs)
- Multi-agent delegation for complex tasks
The Agents API is powered by the open-source Codex harness, giving visibility into coordination logic while OpenAI operates production infrastructure.
Pricing and availability
There are no additional fees for using the Agents API beyond tokens and tools consumed, per OpenAI's pricing page. GPT-Live-1 launched the same week for voice layers at $0.05 per minute, pairable with backend agents.
Constraints to note: US data residency only at beta; Zero Data Retention not supported. Enterprise compliance teams should review before production PHI or regulated data.
Security context you cannot ignore
On September 21, researchers disclosed Heapjack and Overpatch — Codex sandbox escapes fixed in Desktop build 26.818.21641 and CLI 0.149.0. Any developer adopting the Agents API should:
- Run latest harness versions
- Treat third-party repositories as untrusted input
- Monitor MCP tool scopes aggressively
- Prefer OpenAI-hosted sandboxes until your security team validates self-hosted configs
The Agents API does not eliminate sandbox risk — it centralizes responsibility with OpenAI's patching cadence.
Getting started
OpenAI published a quickstart and overview documentation. Typical first projects:
- Internal code review bots with MCP access to GitHub
- Research agents that generate artifacts from structured prompts
- Support triage agents delegating subtasks to specialized sub-agents
Start with read-heavy workflows before granting write or network privileges.
How this fits the stack
The Agents API competes with DIY LangGraph setups, Anthropic's tool-use patterns, and governance layers like WSO2 Agent Manager GA'd September 21. Developers should choose based on operational burden: managed harness versus maximum control.
For learners on Stackademic, the Agents API is the fastest path from "I built a chatbot" to "I built an agent that executes multi-step software tasks" — provided you study the harness and security model, not just the quickstart.
Comments
Loading comments…