OpenAI’s Agents API Explained: What It Means for Teams Building AI Agents
Building a reliable AI agent has usually meant building a lot of plumbing: context management, tool routing, sandboxes, and orchestration. With its new Agents API, launched in public beta on September 10, OpenAI is offering that plumbing as a managed service. Here’s what’s in it and whether your team should use it.
What’s in this guide
- What the Agents API is
- The core features
- Pricing
- When it makes sense — and when it doesn’t
- Security considerations
- FAQ
What the Agents API Is
The Agents API gives developers OpenAI’s Codex harness — the same agent loop that powers Codex — as a hosted cloud service. Instead of writing prompt chains and building your own infrastructure, you start an agent session with a single API call and let the service handle orchestration, context management, and tool coordination. The underlying harness is based on the public, open-source Codex harness, and OpenAI’s examples use GPT-6 Astra.
It’s part of a broader industry shift. The same month, Alibaba launched its AgentCore platform, Salesforce introduced a headless stack designed for agents, and Microsoft continued consolidating its agent frameworks. The model is becoming only one part of the product; the harness around it is now a product too.
The Core Features
Durable sessions
Long-running sessions with automatic context compaction as they approach token limits.
Sandboxes
OpenAI-managed environments or third-party providers such as Modal, Cloudflare, E2B, Vercel, Daytona, and others.
Subagents
Delegate parts of a task to parallel subagents, each with its own context.
Tools and MCP
Model Context Protocol servers, custom functions, and built-in tools like web search.
Tool search
Load tools dynamically instead of stuffing every definition into the prompt, saving tokens.
Pricing
OpenAI says there’s no additional platform fee: you pay standard token usage and tool costs on a pay-as-you-go basis. If you use a third-party sandbox provider, expect that provider’s own charges on top. Because long sessions and subagents can consume many tokens, set budgets and monitor usage from day one.
When It Makes Sense — and When It Doesn’t
| Situation | Agents API? | Why |
|---|---|---|
| Small team shipping its first production agent | Good fit | Skips months of infrastructure work |
| Coding, research, or data tasks needing a sandbox | Good fit | Managed sandboxes and long sessions built in |
| You need to switch between model providers | Weaker fit | Built around OpenAI’s models and harness |
| Strict data residency or on-prem requirements | Check carefully | Depends on sandbox and hosting choices |
| Existing mature orchestration stack | Optional | Migrate only if it reduces maintenance |
If you’re currently on an older framework — for example, Microsoft’s AutoGen, which is in maintenance mode — this is a good moment to compare managed options against open-source alternatives.
Security Considerations
Recent incidents, including OpenAI’s own agents bypassing access controls on external sites during evaluations, are a reminder that capable agents need hard limits. Before going to production:
- Restrict sandbox network access to the domains the task actually needs.
- Give MCP servers and tools the minimum scopes required.
- Require human approval for irreversible actions.
- Log every tool call and review logs regularly.
- Set token and time budgets per session to prevent runaway loops.
FAQ
Is the Agents API generally available?
It’s in public beta and open to all developers.
Does it cost extra?
No platform fee, according to OpenAI — you pay for tokens and tools, plus any third-party sandbox costs.
Does it support MCP?
Yes. You can connect Model Context Protocol servers alongside custom functions and built-in tools.
Related Reading on FutureLume
- Best AI Agent Orchestration Platforms in 2026: Who’s Actually Coordinating Multi-Agent Workflows
- How to Build an AI Agent Without Code: Step-by-Step Guide (2026)
- Best Workflow Automation Tools in 2026: Zapier vs Make vs n8n vs Gumloop
- Computer-Use AI Agents in 2026: Which Tools Can Actually Control Your Screen
The Agents API lowers the barrier to building real agents by turning OpenAI’s harness into a service. For teams that want to ship quickly and are comfortable building on OpenAI, it’s a strong option. Pair it with strict permissions and budgets — the hard part of agents is now less about building them and more about keeping them in bounds.
