OpenAI's Agent Framework & Agents API

In 2026, OpenAI offers two distinct, complementary paradigms for building agentic systems: the self-hosted Agents SDK (pip install openai-agents) and the fully managed OpenAI Agents API (launched in public beta on September 10, 2026 at /v1/agents).

September 10, 2026 Update: The OpenAI Agents API

While the Agents SDK runs inside your own application server and requires you to manage execution loops and local memory, the newly released Agents API provides a fully managed cloud harness:

• Managed Sandboxing: Safely executes code, tools, and computer actions in ephemeral micro-containers without exposing host infrastructure.
• Persistent Session State: Automatically stores multi-turn thread history and checkpoint states on OpenAI's infrastructure.
• Native Context Compaction: Transparently summarizes long-horizon conversations when approaching token limits without breaking tool continuity.
• Parallel Subagent Coordination: Spawns and supervises multiple worker subagents in parallel cloud threads.

Core Primitives (Agents SDK & API)

Primitive	Purpose	Example
Agent	An LLM with instructions + tools	A customer support agent
Handoff	Delegate to another agent	Triage → Billing Agent
Guardrail	Safety validation on input/output	Block PII, reject jailbreaks
Tracing	Observability for debugging	Visualize agent execution flow

from agents import Agent, Runner

# Define a simple agent
support_agent = Agent(
    name="Support Agent",
    instructions="You are a helpful support agent. Answer questions about our product.",
    model="o3-mini"
)

# Run it
result = await Runner.run(support_agent, "How do I reset my password?")
print(result.final_output)

SOE & AOE Operational Runbook for OpenAI Agents

When running enterprise agent fleets, implement top Standard Operating Environment (SOE) and AI Operations Engineering (AOE) controls:

• SOE Standard: Enforce strict organization-level API keys (OPENAI_API_KEY, OPENAI_ORG_ID), pin Python dependencies (openai-agents >= 0.4.0), enforce default developer messages for reasoning models, and configure zero-retention (ZDR) endpoints for enterprise data compliance.
• AOE Architecture: Choose Agents API when you need zero-maintenance infrastructure, auto-scaling subagents, and managed cloud sandboxes; choose Agents SDK when you need strict on-prem/VPC residency, custom vector database stores, or offline tool evaluation. Always route user inputs through the free omni-moderation-latest API before passing to the agent loop.