Back to AI Radar
EmergingAgents & OrchestrationNew entryMarch 2026 New Items

Interesting and early. Worth a spike or exploration session.

OpenAI Agents SDK

The most ergonomic on-ramp if you are already on OpenAI models, but the vendor lock-in risk is real despite claimed multi-provider support.

Agentic·DevTool·Emerging

openai.github.io

Our Take

What It Is

The OpenAI Agents SDK is a lightweight framework for building multi-agent workflows, available in both Python and TypeScript. It's the production-ready successor to Swarm, OpenAI's experimental multi-agent library. The SDK provides very few abstractions by design: agents are LLMs with instructions, tools get automatic schema generation, and handoffs between agents are just tool calls. Built-in guardrails (input, output, and tool-level with PII masking and jailbreak detection) run in parallel with agent execution.

Why It Matters

The SDK's value proposition is simplicity within the OpenAI ecosystem. If your team is already on OpenAI models, this is the lowest-friction way to build agent workflows. Guardrails as first-class primitives (not an afterthought) is a meaningful design choice. The realtime voice agent support is a capability most competing frameworks lack entirely. And the Frontier enterprise platform, launched February 2026, provides a deployment and management layer on top of the SDK.

Key Developments

  • Mar 2026: v0.12.1 released. Realtime API migrated to gpt-realtime model (GA). WebSocket mode for Responses API added as opt-in.
  • Feb 2026: OpenAI launched Frontier, an enterprise platform for building, deploying, and managing AI agents.
  • 2025-2026: TypeScript SDK released. Sessions for persistent memory, human-in-the-loop, and tool guardrails added.
  • Mar 2025: Initially launched as production-ready successor to Swarm.

What to Watch

The vendor lock-in question is structural. While the SDK technically works with 100+ LLMs via Chat Completions API, the Responses API is designed to prevent provider switching. The Frontier platform tightens coupling between agent orchestration and OpenAI's infrastructure. At v0.12.1, the API surface is not yet stable, so watch for breaking changes. And compare the durable execution story against PydanticAI, which handles progress preservation more maturely.

Strengths

  • Low abstraction overhead: Very few abstractions by design. Agents are LLMs with instructions, tools get automatic schema generation, handoffs are tool calls.
  • Guardrails are first-class: Input, output, and tool guardrails run in parallel with execution. PII masking and jailbreak detection built in.
  • Tracing built in: Visualising, debugging, and monitoring agent workflows without a separate observability stack.
  • Realtime voice agents: Native support for voice agents with interruption detection and context management. Most competitors lack this entirely.

Considerations

  • Vendor lock-in is structural: Responses API designed to prevent provider switching via base_url changes. Optimised for OpenAI models.
  • Still pre-1.0: At v0.12.1, the API surface is not stable. Breaking changes between minor versions have occurred.
  • Memory and persistence gaps: Long-term memory needs separate layering. Durable execution across restarts is less mature than PydanticAI's approach.
  • Enterprise lock-in: Frontier platform tightens coupling to OpenAI's infrastructure. API price hikes or regional restrictions hit your product directly.

Resources

More in Agents & Orchestration

OpenAI Agents SDK· A2A Protocol· PydanticAI· AI Browser Use· Agentic RAG· CrewAI· Multi-agent Orchestration· OpenClaw· Chain-of-Thought· LangGraph· Model Context Protocol· Tool Use / Function Calling

Back to AI Radar