KDCube
← AI Industry News

OpenAI's Agents API Hosts the Whole Agent Stack

OpenAI's new Agents API (public beta, Sept 10) puts the Codex harness behind one call — managed subagents, sandboxes, MCP, and multi-day sessions — while a same-day Data Agent reaches into Snowflake, BigQuery, and Databricks. Meanwhile Zscaler's Agentic SOC turns agents on attackers and AWS Bedrock AgentCore adds cross...

Highlights

  • OpenAI put the Codex harness behind one API call: the new Agents API (public beta, Sept 10) runs managed cloud agents with subagents, sandboxes, MCP, and multi-day sessions — collapsing the orchestration layer teams used to hand-build. (OpenAI)
  • Same day, OpenAI shipped a Data Agent in ChatGPT for Work that connects to Snowflake, BigQuery, Databricks, and Redshift and writes to Tableau/Power BI — natural-language analytics without a query editor. (Unite.AI)
  • Zscaler Agentic SOC (Sept 9) turns the other direction: expert-validated AI agents plus a decoy mesh to triage, verdict, and contain AI-driven attacks at machine speed, running on Anthropic and OpenAI models. (Help Net Security)
  • AWS answered by making Bedrock AgentCore Evaluations score TypeScript agents built on LangGraph, OpenAI Agents, and Strands — and added a user Consent Portal to Identity. (AWS)

Key Signals

  1. OpenAI's Agents API kills the orchestration layer Sept 10, 2026

    The Agents API exposes the same open-source Codex harness that powers Codex and ChatGPT for Work: it manages context compaction, tool use, recovery, and subagent coordination while your app supplies tools and picks the execution environment. Agents can run in a sandbox to execute code, edit files, hit MCP servers, and produce artifacts across sessions that stay alive for days. (OpenAI, MarkTechPost) For teams who wired their own graphs in LangGraph or CrewAI, the calculus shifts toward hosted convenience — and toward OpenAI-managed lock-in.

  2. The connector war moves to the data layer Sept 10, 2026

    OpenAI's Data Agent reaches into approved warehouses (Snowflake, BigQuery, Databricks, Redshift, ClickHouse, MongoDB) and BI surfaces (Tableau, Power BI, Sigma, ThoughtSpot), investigating metric changes and building shareable dashboards from a single conversation. (Unite.AI) This is RAG-over-governed-data as a product feature, and it puts OpenAI directly against the semantic-layer and BI incumbents.

  3. Defenders go agentic too Sept 9, 2026

    Zscaler Agentic SOC pairs inline zero-trust telemetry with a global decoy mesh and specialized agents for triage, root-cause investigation, verdicting, and automated containment — explicitly framed as a response to attackers shifting from prompts to automated agentic chains. (Help Net Security) The stack now has agents on both sides of the wire.

Why It Matters / What To Watch

  1. Managed runtime vs. portable control
    • Weigh the Agents API's hosted harness against portability: OpenAI owns session state, orchestration, and recovery, so plan an exit path before betting a fleet on it. (OpenAI)
    • Watch cross-runtime evaluation as the hedge — Bedrock AgentCore now evaluates OpenAI Agents, LangGraph, and Strands agents in TypeScript, so you can benchmark before committing. (AWS)
  2. Data and consent are the new attack surface
    • If you deploy the Data Agent, gate warehouse connectors behind least-privilege roles — natural-language access to Snowflake and Databricks is only as safe as the credentials behind it. (Unite.AI)
    • Track consent primitives: AgentCore Identity's new Consent Portal lets end users approve what an agent may touch on their behalf, a pattern governance teams should demand everywhere. (AWS)
  3. Inspect third-party agent components before they run
    • As agent/skill/MCP registries fill up, pre-deployment screening matters — Tenable's CyberAgents Exchange AI Inspector combines GPT cyber models, Tenable One exposure analysis, and human review to vet skills and MCP servers before production. (Tenable)

Quick Links