- OpenAI shipped a major Agents SDK update on April 15 introducing native sandboxing, a long-horizon orchestration harness, subagents, code mode, and provider-agnostic support for 100+ LLMs.
- The new two-layer architecture separates the harness (instructions, tools, approvals, tracing) from the sandbox (filesystem, shell, memory), enabling isolated and durable agent execution.
- Google shipped Gemini CLI subagents the same day, allowing specialized expert agents to run in parallel with dedicated context windows, tools, and MCP servers.
- Both releases signal a clear industry convergence toward multi-agent, sandbox-first developer tooling — the era of single-prompt AI coding assistants is giving way to orchestrated agent fleets.
Within 24 hours of each other, OpenAI and Google shipped what may be the most consequential developer-tool updates of 2026 so far. OpenAI’s Agents SDK now lets enterprises spin up sandboxed, crash-recoverable agent sessions that mount cloud storage and respect Unix-level permissions, while Google’s Gemini CLI gained a parallel subagent architecture that keeps primary sessions lean. Together, the two releases crystallize a new consensus: production AI agents need isolation, orchestration, and multi-model flexibility — not just a bigger context window.
OpenAI Agents SDK: From Prototype to Production
The Two-Layer Architecture
The core design change is a clean separation between the harness and the sandbox. The harness handles instructions, tool definitions, approval workflows, tracing, handoffs, and resume bookkeeping. The sandbox provides filesystem tools, shell access, a skill and memory system, compaction, and workspace customization through a new Manifest file. These two layers can run on the same machine for local development or on separate infrastructure for enterprise-grade isolation.
Manifest and Cloud-Native Storage
The Manifest is a portable configuration layer that stages local files, clones Git repositories, creates output directories, and mounts external storage from S3, GCS, Azure Blob Storage, or Cloudflare R2. Combined with Unix-level permissions — for example, a read-only dataroom with a write-enabled output directory — the system gives compliance-heavy industries like finance and healthcare a credible path to deploying autonomous agents without surrendering access control.
Durability: Snapshots and Rehydration
Long-running agents inevitably crash. The SDK now supports built-in snapshotting and rehydration, allowing a session to be restored in a fresh sandbox and continue from its last saved state. For tasks that span hours — large-scale code migrations, dataset transformations, multi-repo refactors — this eliminates the fragility that plagued earlier agent frameworks and brings genuine durability to agentic workflows.
Provider-Agnostic by Default
Perhaps the most strategically revealing detail: the SDK now supports 100+ LLMs out of the box. OpenAI is explicitly positioning the Agents SDK not as a lock-in vehicle for GPT models, but as a universal orchestration layer. Developers can swap models per agent, per task, or per cost tier — a direct response to the growing multi-model reality where teams use different providers for different workloads.
AI Biz Insider Analysis — OpenAI is making a calculated bet: own the orchestration layer even if you don’t own every model in the pipeline. By open-sourcing the SDK with provider-agnostic support and enterprise-grade sandboxing, OpenAI is trying to become the Kubernetes of agentic AI — the control plane that everyone builds on, regardless of which inference engine runs underneath. If it works, revenue shifts from per-token charges to platform fees and enterprise support contracts.
Gemini CLI: Subagents Go Mainstream
Isolated Context, Parallel Execution
Released on April 15, Google’s Gemini CLI now supports subagents — specialized expert agents that operate within their own separate context windows, custom system instructions, and curated tool sets. Each subagent’s multi-step execution consolidates into a single summary returned to the main agent, preventing the context degradation that plagues long coding sessions. Users invoke them with @agent notation: @frontend-specialist, @codebase_investigator, or any custom agent defined in a Markdown file with YAML frontmatter.
Built-In and Custom Agents
Gemini CLI ships with three built-in subagents: a generalist with full tool access, a cli_help agent for answering Gemini CLI feature questions, and a codebase_investigator for mapping dependencies and exploring code structure. Developers can define custom subagents at the global level (~/.gemini/agents) or per-project (.gemini/agents), and extensions can bundle agents in an agents/ directory — creating a composable, shareable agent ecosystem.
AI Biz Insider Analysis — Google’s approach is complementary but different. Where OpenAI focuses on infrastructure-level sandboxing and cloud-native durability, Google optimizes for developer experience at the CLI level. The @agent notation and Markdown-based configuration lower the barrier to multi-agent setups dramatically. Expect Anthropic’s Claude Code — which already supports sub-agents via the Agent tool — to respond with its own enhancements soon, intensifying the three-way platform race for developer mindshare.
What This Means for Enterprise Teams
The simultaneous releases mark an inflection point. Enterprise engineering teams evaluating agentic tooling now face a clearer taxonomy: OpenAI’s SDK targets backend infrastructure teams that need cloud-mounted, permission-controlled, crash-recoverable agent pipelines; Google’s Gemini CLI targets individual developers and small teams that want parallel expertise without context overflow. Both are betting that multi-agent orchestration — not monolithic chat — is the future of AI-assisted software development.
For CTOs and engineering leads, the practical takeaway is straightforward: if your agents need to touch production data, mount storage, and survive crashes, the OpenAI Agents SDK’s sandbox-plus-harness model deserves evaluation. If your priority is developer velocity on local codebases with specialized expertise, Gemini CLI subagents offer the fastest path to multi-agent workflows today. And with MCP crossing 97 million installs and the Agentic AI Foundation formalizing interoperability standards, the walls between these platforms are getting thinner by the quarter.
Related
- Anthropic Shifts to Usage-Based Enterprise Billing as ARR Triples to $30B
- Google Launches Native Gemini App for Mac
- Project Glasswing: Claude Mythos Preview and the $100M Cybersecurity Alliance
- TOP3 DIGEST — Anthropic Opus 4.7, OpenAI Trusted Access, AI and the Human 20%
Sources
- OpenAI — The Next Evolution of the Agents SDK (April 15, 2026)
- Google Developers Blog — Subagents Have Arrived in Gemini CLI (April 15, 2026)
AI Biz Insider · AI Trends EN · aibizinsider.com
댓글 남기기