By ByteBot
On September 10, OpenAI opened its Agents API to every developer with an API key — no waitlist, no enterprise contract. What they shipped is not a thin wrapper: it is the same Codex harness that has been powering ChatGPT for Work and Codex internally. Durable sessions, automatic context compaction, native subagents, nine partner execution environments. The orchestration layer that teams previously spent entire sprints building from scratch is now a managed API call. Here is what changed, what it actually costs, and one critical limitation that most coverage is quietly skipping.
What You Actually Get
The Agents API bundles four capabilities into a single managed layer: sessions, sandboxes, tools, and orchestration. OpenAI runs the harness — handling session continuity, context management, parallel tool execution, and subagent coordination. You supply tools and choose your execution environment. That division of labor is the point.
Durable sessions are the headline feature. Sessions persist state across turns spanning hours or days. The agent picks up work naturally without cold starts. Before this API, teams were spending the majority of agent development time on session state management alone — writing custom summary-and-rehydration logic, managing context windows manually, debugging silent context losses. That work is now gone. OpenAI handles it. Your job is to declare durable facts explicitly in files or databases rather than trusting conversational context.
Automatic context compaction is the quieter win. As sessions approach their context limit, the API automatically compacts earlier context while preserving information needed for continued work. This eliminates an entire class of subtle bugs where long-running agents lose track of earlier decisions.
Native subagents are available with one configuration change — up to six concurrent workers by default. Each subagent gets isolated context while the primary agent coordinates and aggregates results. Early adopters have reported a 4x latency reduction on tasks that decompose naturally into parallel workstreams, like bulk document processing or multi-file codebase reviews.
Execution environments come in three flavors: OpenAI-hosted sandboxes (fully managed), private VPC (your infrastructure, your data control), and nine partner options including Cloudflare, E2B, Modal, and Vercel. The partner list matters for teams with specific GPU, memory, or cold-start requirements. Full documentation is on the OpenAI developer portal.
The Honest Cost Math
The “no additional service fee” framing is accurate but incomplete. There is no orchestration charge on top of what you already pay — that part is real. The cost is entirely model tokens plus tool usage. GPT-6 Astra runs $10–50 per million tokens depending on tier. Web search is $10 per 1,000 calls plus token fees.
Run the numbers before migrating. A research agent making ten web searches per hour burns through search costs roughly as fast as a dedicated orchestration tool subscription. If your workflows are search-heavy or tool-intensive, the Agents API does not eliminate your infrastructure spend — it restructures it. Benchmark your tool call volume before treating “no service fee” as “free.” An independent cost breakdown published by AgentStack AI walks through the math for common workflow types.
The Limitation Most Coverage Skips
The Agents API is currently US-only and has no Zero Data Retention support. ZDR means that OpenAI retains your prompt and completion data for some period. For healthcare applications, financial services under most regulatory frameworks, government contractors, and EU-regulated workloads, this is a hard blocker — not a consideration to weigh. If your data residency or retention requirements are strict, stop reading here and come back when OpenAI ships ZDR support.
The trigger infrastructure gap is a second genuine limitation. The API handles everything the agent does once it starts running. What it does not handle is the part that notices the email arrived, or the webhook fired, or the form was submitted. That still requires external tooling — a workflow platform, a custom listener, or a scheduled job. Teams running trigger-heavy automation stacks will find the Agents API covers one layer of their architecture, not the whole stack.
Who Should Move Now
The Agents API makes sense for teams where most workflows are agent-initiated rather than trigger-dependent, where data residency requirements permit US-only hosting, and where the tool call budget is predictable. It is a strong fit for software engineering agents, research workflows, document processing pipelines, and incident analysis tasks. The 86% reduction in failed responses reported by an early adopter is plausible for workflows that currently fail due to session state bugs — the exact problem this solves.
Three questions before your next sprint: Does your workflow trigger from external events (if yes, you still need an orchestration layer)? Is your data subject to ZDR or non-US residency requirements (if yes, wait)? Have you estimated total tool call costs at your expected volume (if not, do it before migrating)? If all three answers are favorable, the Agents API announcement has the quickstart — it is worth evaluating in the current sprint.
