
Tech • IA • Crypto
OpenAI’s updated Agents SDK introduces a Codex-style harness, sandboxed execution, and persistent state features to make long-running, production-grade AI agents easier to build and deploy.
AI models are increasingly capable of handling tasks over extended periods, from minutes to days. Internal tools have demonstrated agents running continuously for up to a week, completing complex workflows such as coding, data analysis, and security scanning. This shift is driving demand for infrastructure that supports sustained, autonomous operation.
Despite improved model capabilities, deploying agents in real-world systems presents challenges. Developers must balance performance with flexibility across model providers, manage state across failures, and handle orchestration logic. Issues like container crashes, state loss, and secure handling of secrets complicate scaling.
The SDK now integrates a Codex-inspired harness that automates the agent loop, including tool use, context updates, and task continuation. Features like asynchronous shell execution, command tracking, and automatic context compaction allow agents to operate continuously without manual orchestration.
A key architectural change splits the agent harness from the execution environment. Instead of coupling logic and compute in a single container, agents can run in ephemeral sandboxes while orchestration persists elsewhere. This enables recovery from failures and avoids state loss when containers terminate.
Agents can operate inside isolated sandbox environments with access to files, enabling tasks like code editing, document processing, and asset generation. These sandboxes can run locally via Docker or in cloud platforms such as Modal, Cloudflare, Vercel, and others, offering flexibility for development and production.
The SDK introduces built-in state management by snapshotting both the file system and conversation history. These snapshots can be stored locally or in cloud storage like R2 and later reloaded, allowing agents to resume tasks seamlessly even after interruptions.
A new Skills API allows developers to package domain-specific logic, instructions, and resources into reusable units. Skills can be versioned, stored centrally, and loaded dynamically, enabling agents to perform specialized tasks such as tax preparation or content editing with consistent behavior.
The Responses API now includes a hosted shell tool that spins up temporary containers for single tasks. Developers can upload files, execute code, and retrieve outputs in one call, offering a lightweight alternative to full agent deployment.
Developers can extend agents with custom tools, defined as functions that the model can call automatically. These tools support validation, guardrails, timeouts, and dynamic enablement, enabling integration with external systems such as task trackers or databases.
The SDK introduces a manifest system to define file system structure and data sources. Files can be copied into sandboxes or mounted from external storage like S3-compatible buckets, balancing performance with scalability depending on workload needs.
While still evolving, the framework supports orchestrating multiple agents working in parallel. Coordination can occur through shared storage, messaging, or supervisory agents, with expectations that large-scale multi-agent systems will become more common.
The updated Agents SDK reflects a shift toward persistent, autonomous AI systems by combining structured orchestration, sandboxed execution, and built-in state management into a flexible developer framework.