OpenAI released three major infrastructure products for AI agent development on September 22, 2026, representing a significant consolidation of the fragmented tooling landscape developers have faced when building autonomous systems.
The Three Pillars
The company announced the Agents SDK v2, AgentKit, and the Agents API—each targeting different deployment patterns but sharing common underlying infrastructure. This coordinated release suggests OpenAI is attempting to own the full stack of agent development, from local prototyping to enterprise deployment.
Agents SDK: Native Sandbox Execution
The updated Agents SDK introduces what OpenAI calls a "model-native harness"—standardized infrastructure designed specifically for how OpenAI's frontier models operate. The key addition is native sandbox execution, allowing agents to run in controlled environments with file system access, code execution, and dependency management.
Developers can now configure sandboxes through a Manifest abstraction that supports multiple providers including Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel. The SDK separates the agent harness from compute execution, which OpenAI claims improves security by keeping credentials out of environments where model-generated code runs.
The SDK also introduces features for long-running tasks: configurable memory, sandbox-aware orchestration, and built-in snapshotting with rehydration—allowing agents to resume from checkpoints if containers fail or expire.
AgentKit: Enterprise Visual Builder
AgentKit targets enterprise teams with a visual canvas for building multi-agent workflows. Three components define the platform:
- Agent Builder: A drag-and-drop interface for creating and versioning agent workflows, with support for guardrails, preview runs, and inline evaluation
- Connector Registry: A central admin panel for managing data connections across Dropbox, Google Drive, SharePoint, Microsoft Teams, and third-party MCP servers
- ChatKit: An embeddable chat interface toolkit for deploying conversational agents
OpenAI cited customer results: Ramp reported going from "blank canvas to buyer agent in just a few hours" with "70% slash in iteration cycles," while LY Corporation built a work assistant in under two hours.
Notably, OpenAI included a deprecation notice: Agent Builder and Evals products will shut down on November 30, 2026, with the company directing users to the Agents SDK for code-based workflows and Workspace Agents in ChatGPT for natural language prompting.
Agents API: Managed Cloud Service
The Agents API offers the same Codex harness and infrastructure that powers OpenAI's own Codex and ChatGPT for Work products, but delivered as a fully managed cloud service. Developers can create production-ready agents with a single API call specifying model, tools, environment, and task delegation preferences.
The API supports multiple deployment patterns: OpenAI-hosted sandboxes, self-hosted infrastructure, or integration with sandbox partners (adding DigitalOcean and Oracle to the SDK's provider list). Built-in capabilities include automatic context compaction for long sessions, tool search to reduce token usage, programmatic parallel tool calling, and multi-agent support with up to 3 concurrent subagents.
Customer claims from the announcement include evaluation scores improving from 0.71 to 0.85 with 4x latency reduction (Ciridae), 60% cost reduction per case (SafetyKit), and 86% reduction in failed agent responses from harness-sandbox separation (Hypha).
Architecture Decisions
All three products build on what OpenAI describes as a "model-native harness"—infrastructure aligned with how frontier models perform best, rather than generic orchestration frameworks. The company explicitly positioned this against "model-agnostic frameworks" that sacrifice model-specific optimizations for flexibility.
The separation of harness from compute appears consistently across products, suggesting this is a core architectural bet. OpenAI states this enables durable execution (state externalization for recovery), improved security (credential isolation), and scalability (parallel sandbox invocation).
Competitive Context
These releases arrive as cloud infrastructure for agents reaches maturity. According to independent analysis from Forkast, AWS, Google, and Microsoft Azure now all offer production-grade agent runtimes with service level agreements—shifting competition from availability to architectural efficiency.
OpenAI's approach differs from cloud providers by owning the full vertical: model, harness, and optional infrastructure. This creates distinct risks and benefits for developers—tighter integration and potential performance gains versus vendor lock-in and reduced deployment flexibility.
Limitations and Open Questions
Several important limitations are not addressed in the announcements:
- Pricing transparency: While OpenAI states "standard API pricing" applies, actual costs for sandbox execution, subagent orchestration, and context compaction are unspecified
- Availability gaps: Python SDK launched first, with TypeScript "planned for a future release"
- Deprecation timeline: The November 2026 shutdown of Agent Builder and Evals gives enterprises limited migration time
- Model lock-in: The "model-native" positioning implies optimizations may not transfer to other providers
The customer testimonials cite impressive results, but without independent verification or disclosed methodology, these should be treated as promotional claims rather than established benchmarks.
What Must Be Verified
Claims requiring additional scrutiny include:
- Specific cold-start performance numbers for OpenAI-hosted sandboxes (not provided in announcements)
- Subagent latency reduction quantification beyond cited customer testimonials
- Cost comparisons to self-hosted alternatives at scale
- Security properties of the harness-compute separation under adversarial testing
Evaluating OpenAI's Infrastructure Bet
OpenAI's coordinated release represents a significant consolidation attempt in agent infrastructure. By offering SDK, visual builder, and managed API options sharing the same underlying harness, the company is positioning itself as the default infrastructure layer for agent development—regardless of where the actual compute runs.
The approach contrasts with the Model Context Protocol's (MCP) recent stateless specification revision, which emphasizes interoperability and load-balancer-friendly design. OpenAI's products support MCP as one tool option among many, but the overall architecture suggests a preference for vertically integrated solutions.
Whether this consolidation benefits developers depends on whether OpenAI's performance claims hold under independent evaluation, and whether the convenience of integrated infrastructure outweighs the risks of single-vendor dependence.