OpenAI's Agents SDK now opens the file system, the shell, and the editor
Codex-style agents acquire the keys to the computer; pay-as-you-go pricing follows the same week.
Agents can now inspect files, run shell commands, and edit code inside sandboxes — the long-horizon autonomyThe ability of an artificial intelligence system to pursue complex, multi-step goals over an extended period without requiring human intervention or prompting at each intermediate step. that was previously the province of research preview. Priced per successful action, not per minute, the new Agents SDK arrives with a commercial model designed to align vendor revenue with customer outcomes, and a risk surface expanded accordingly.
The developer-tooling question of the past eighteen months has been narrow and specific: when do agents acquire genuine computer-use capability, and who ships it first at production quality. OpenAI's answer, delivered with unusually restrained marketing copy, is now. The SDK exposes file-system, shell, and editor primitives through a unified tool-call interfaceA programming boundary that allows an AI model to execute external functions—like reading a file or running a script—by formatting its output as a structured command rather than conversational text., with sandbox boundaries enforced at the process and network level and policies configurable per organisation.
The detail that matters commercially is the pricing mechanism. According to OpenAI's own documentation, successful-action billing is determined by a completion signal the agent emits and the customer's environment confirms — a design that, on paper, transfers the cost of retries and failed attempts from the buyer to the vendor. Sources familiar with early enterprise pilots describe pricing in the low single-digit dollars per completed coding task, with volume tiers that have not yet been publicly disclosed. The model aligns incentives in a way per-minute pricing never did; it also creates a new class of gaming risk around what counts as success.
The winners are the developer-productivity teams inside large software organisations, who have spent a year building scaffolding for exactly this capability and now have a first-party option. The losers are the agent-framework startups whose value proposition was the glue OpenAI has now absorbed into its own SDK, and the per-seat coding-assistant incumbents whose pricing model suddenly looks like a legacy abstraction.
What the release opens is a category — outcome-priced agentic softwareSoftware designed to pursue open-ended goals by planning intermediate steps and executing them autonomously, rather than following a rigid set of pre-programmed rules. — that every enterprise buyer is about to demand from every vendor. What it forecloses is the comfortable interim in which agents were billed by consumption rather than by result. The SDK will be copied within the quarter. The pricing model will take longer, because it requires infrastructure most vendors do not yet have.