OpenAI released a significant update to its Agents Software Development Kit (SDK) on April 15, introducing sandboxing capabilities and a long-horizon task harness designed to make autonomous AI agents safer and more practical for enterprise deployment. The update marks the second major expansion of the SDK since its launch and arrives amid mounting concerns from IT leaders about the operational risks of unsupervised agent behavior.
According to TechCrunch, the new features are available to all customers via the OpenAI API at standard pricing, with Python support launching first and TypeScript support planned for later in 2026.
What's New in the Update
Sandboxing: The SDK now offers sandboxed execution environments, allowing agents to operate in controlled computer environments rather than directly on enterprise infrastructure. Enterprises can either bring their own sandbox providers or use built-in support for seven partner platforms: Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel.
Long-Horizon Harness: Paired with sandboxing, the harness enables agents to execute multi-step, complex tasks that span longer time horizons. Previously, agents were typically limited to relatively short reasoning chains before losing coherence or requiring human intervention.
Upcoming Capabilities: OpenAI confirmed it is working to bring additional features — including "code mode" and subagent orchestration — to both Python and TypeScript in future releases.
Why Sandboxing Matters Now
The sandboxing push addresses a real operational risk. OutSystems research published earlier this year found that 94% of enterprise leaders raised concerns about "agent sprawl" — the proliferation of unmanaged, unmonitored AI agents across business functions.
The same study noted that while 97% of executives say their companies have deployed AI agents in the past year and 52% of employees already use them, only 23% of organizations report significant ROI from agent deployments. A major obstacle: containing agent behavior within safe operational boundaries.
Sandboxing directly addresses this. By isolating agent execution into controlled environments, enterprises reduce blast radius when an agent goes off-script — a frequent failure mode with long-running autonomous tasks.
Enterprise Revenue Driving the Investment
The update comes as enterprise now accounts for more than 40% of OpenAI's total revenue, with the company stating it is on track to reach revenue parity between enterprise and consumer segments by the end of 2026. Enterprise customers represent a dramatically different buyer profile than OpenAI's initial consumer-focused user base — one where controls, auditability, and isolation are non-negotiable purchase criteria.
Writer's 2026 enterprise AI adoption survey found that 79% of organizations face challenges adopting AI — a double-digit jump from 2025 — and 86% said their AI budgets will increase this year, with nearly 40% planning increases of 10% or more.
Partner Ecosystem Signals Infrastructure Maturity
The seven launch partners — Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel — represent a cross-section of serverless compute, developer infrastructure, and isolated execution providers. Their inclusion signals that the AI agent stack is consolidating around a set of infrastructure primitives similar to what cloud computing underwent in 2015-2020.
Cloudflare's participation is particularly notable given its existing Workers platform serves millions of developers building edge-deployed applications. For enterprises already using Cloudflare for other workloads, the native integration removes a common friction point: getting procurement approval for yet another vendor.
Market Context: The Multi-Agent Breakout
Both Forrester and Gartner have identified 2026 as the breakthrough year for multi-agent systems, where specialized agents coordinate under central orchestration. Gartner predicts 40% of enterprise applications will include task-specific AI agents by the end of this year.
The OpenAI update aligns with this thesis. The combination of sandboxing plus a long-horizon harness creates the foundation for reliable subagent orchestration — where a parent agent delegates specific tasks to specialized subagents running in isolated environments.
How This Affects Virtual Assistant Workflows
For businesses using virtual assistants and the tools VAs operate, these updates matter in several ways:
- Safer automation expansion: Sandboxing makes it easier for VAs to deploy agent-powered workflows without risking broader system access.
- Longer task chains: The harness enables end-to-end automation of processes that previously required human handoffs — lead research, document processing, multi-channel outreach.
- Tool compatibility: Popular VA workflow tools built on the OpenAI SDK will inherit these improvements automatically as providers update.
However, the technology also raises the bar on VA capabilities. As AI agents take over more repetitive work, human VAs are moving toward judgment, orchestration, and client relationship tasks — reinforcing the trend toward specialized virtual assistant services with deep domain expertise.
The Bigger Picture
OpenAI's update is part of a broader industry shift from AI experimentation to production-grade deployment. The features released April 15 — sandboxes, harnesses, partner integrations — are the unglamorous infrastructure that enterprise buyers actually care about.
For VA businesses, the takeaway is simple: the tools available to virtual assistants are rapidly becoming more capable. The most valuable VAs in 2026 are those who know when to deploy agent automation, when to intervene, and how to audit outcomes — not the ones executing routine tasks manually.
Sources: