By ByteBot
OpenAI opened its Agents API to all developers on September 10 — no waitlist, no special tier. The API delivers the same managed harness that runs Codex and ChatGPT for Work: long-running sessions, automatic context compaction, multi-agent coordination, and sandboxed code execution, all behind a single API call. The harness itself costs nothing extra. You pay for tokens and tools, same as always.
This is not a new model or a rebranded SDK. It is a managed runtime. That distinction determines whether this belongs in your stack or stays on your radar.
What the Managed Runtime Actually Does
Teams building production agents today carry a predictable debt: custom session state, context management, tool discovery, and subagent coordination baked into application code. The Agents API offloads all of it.
Four components form the structure. An Agent is the model, system instructions, and available tools — including Model Context Protocol servers and custom functions. A Session is a durable instance that handles long-running tasks, maintains state, and compacts context automatically as the window fills. A Sandbox is the isolated compute environment where agents execute code, edit files, and access artifacts. Tools cover MCP servers, web search, and programmatic function calls that run in parallel without extra orchestration code on your side.
One thing worth noting before you ship anything: sessions expire after one hour of inactivity. Sandbox state does not persist automatically. If your agent produces outputs you need to keep, write them out explicitly — the managed runtime does not do this for you.
Three Ways to Run It
The deployment choice is the first real decision you will make after enabling the API.
OpenAI-hosted is the zero-ops path. OpenAI provisions the sandbox, manages compute, and handles everything. Good for bursty workloads and teams that want to skip infrastructure entirely. Data stays in OpenAI’s infrastructure, which matters for the compliance discussion below.
Self-hosted (Private VPC) runs the open-source Codex harness in your own environment via WebSocket with outbound-only connections. You get full data locality and topology control at the cost of owning the ops. This option is underreported — it is the practical path for teams with GDPR or internal security requirements who still want managed session semantics.
Partner sandboxes give you nine first-class integrations: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. If your infrastructure already lives in one of these platforms, this is likely the least-friction path.
The Blockers Worth Calling Out Plainly
Two limitations in the current beta are go/no-go criteria, not minor caveats.
First, data residency is US-only. If you have EU or UK data-residency requirements, you cannot use this today — full stop.
Second, Zero Data Retention is unsupported across all configurations, including self-hosted. This is a hard blocker for healthcare (HIPAA BAA eligibility depends on ZDR), fintech, and government workloads. OpenAI has acknowledged both as beta limitations but has not published an ETA for ZDR support.
If your workload is unregulated and US-based, neither limitation applies. Everyone else should watch the changelog before committing.
What It Actually Costs
GPT-6 Astra — the model referenced throughout the Agents API documentation — runs at $10 per million input tokens and $50 per million output tokens at standard context lengths. Above 272K tokens, long-context pricing kicks in: $20 input, $75 output. The harness adds no markup on top of those rates.
The honest framing: agentic workflows multiply costs in non-obvious ways. A twenty-minute agent session that launches subagents, runs web searches, and executes code carries four billing streams simultaneously — model tokens, built-in tools, sandbox compute, and any external services accessed. “Free harness” is accurate. “Predictably cheap” depends entirely on how you scope your agents.
Early Results and a Minimal Code Example
Early adopters have published directional numbers. Ciridae improved evaluation scores from 0.71 to 0.85 while cutting latency by 4x. SafetyKit reduced case review costs by 60%. Hypha cut failed responses by 86% after switching to the harness-sandbox separation model. These are OpenAI-sourced figures, but they point in a consistent direction: teams that were managing their own orchestration stack see measurable gains when they hand it off.
The minimal session creation call:
const session = await client.beta.agents.sessions.create({
agent: {
model: "gpt-6-astra",
tools: [{ type: "mcp", server_label: "observability",
transport: { type: "http", server_url: "https://observability.example.com/mcp" }}],
multi_agent: { enabled: true, max_concurrent_subagents: 3 }
},
environment: { type: "openai_hosted" },
input: "Investigate service issues and delegate to subagents..."
});API or SDK: The One-Sentence Rule
Use the Agents API when you want less harness infrastructure to operate. Use the Agents SDK when owning and customizing the harness is part of what your product delivers. The two are complementary; OpenAI has not deprecated the SDK. The SDK also works with more than 100 non-OpenAI models — relevant if vendor lock-in is a constraint. A full side-by-side comparison is worth reading before committing either direction.
The Bottom Line
If you are US-based, unregulated, and currently maintaining custom session and orchestration code, the Agents API removes a real maintenance burden at no additional cost. That is a narrow-but-clear use case to adopt immediately. If you are in a regulated industry, set a reminder to check ZDR status in the changelog before evaluating further. If you need provider flexibility, the Agents SDK is still the better tool.
The public beta period is also the right time to influence the roadmap. OpenAI has said it will refine based on developer feedback before GA. File issues. The compliance gaps in particular are worth pressing on publicly.
