KDCube
← AI Industry News

OpenAI's Agents API Ships as a Managed Runtime at DevDay 2026

OpenAI's DevDay keynote turned the Agents API into a public-beta managed runtime, with OpenAI handling sessions, orchestration, and context compaction while builders bring tools — plus a co-announced Bedrock Managed Agents that keeps data inside AWS. It also unveiled always-on Dots agents on GPT-6 Astra and a ...

Lead signals

Highlights

  • OpenAI's DevDay keynote (Sept 29) shipped the Agents API as a public-beta managed runtime: OpenAI handles sessions, orchestration, context compaction, and recovery while builders supply tools and execution environments — the "code hints" from the prior issue are now a real product. (runtimewire)
  • OpenAI and Amazon jointly introduced Bedrock Managed Agents, running the same harness where "agents run in the customer's environment, inference runs on Bedrock, and data stays within AWS." (runtimewire)
  • Dots, always-on agents powered by GPT-6 Astra, each get their own cloud computer and browser, run 24/7, and reach 4,000+ apps — gated to Pro and Business Premium. (Decrypt)
  • New GPT-6.1 Sol claims near-Astra coding/computer-use at one-fifth the token price ($2 / $0.10 cached / $10 per M), alongside a $500/month Pro 500 tier. (Pulse 2)
What changed

Key Signals

  1. The Agents API turns orchestration into a hosted serviceDevDay, Sept 29 2026

    The runtime lets agents execute code, edit files, connect to MCP servers, delegate to subagents, and drive an OpenAI-hosted browser for computer use, with durable session state and monitorable session events. For teams running many agentic apps, context compaction and recovery move off your stack — less ops burden, less low-level control. (runtimewire, OpenAI)

  2. Bedrock Managed Agents is the governed-deployment templateDevDay, Sept 29 2026

    Co-announced with AWS, it pairs OpenAI's agent harness with data residency: inference on Bedrock, agent execution and data inside the customer's AWS boundary. Expect Vertex AI Agent Builder and Copilot Studio to answer with comparable "your cloud, your data" managed-agent framing. (runtimewire)

  3. Economics shift again with GPT-6.1 SolDevDay, Sept 29 2026

    Sol at roughly 1/5 of Astra's token prices reframes the agent-fleet math that dominated recent issues (Anthropic's 950-agent run). Meanwhile Pro 500 at $500/month — 25× the Plus allowance and the only path to the faster Ultrafast tier (up to 8× in Codex, 6× in the API) — signals aggressive premium tiering for heavy agent users. (Pulse 2)

Operator playbook

Why It Matters / What To Watch

  1. Managed runtime vs. framework control is now a live build decision.
    • If you adopt the Agents API, plan for OpenAI owning compaction and recovery — audit what session events you can actually observe before betting governance on it. (runtimewire)
    • Watch the Agent Builder / Evals sunset on Nov 30, 2026: OpenAI steers code-first workflows to the Agents SDK and prompt-first ones to Workspace Agents, so migrate now rather than at the deadline. (OpenAI)
  2. Portability hedges against managed lock-in.
    • LangChain 1.4.0's first-party MCPAdapter (langchain.mcp) turns any MCP server — HTTP, local, or in-process FastMCP — into create_agent tools, keeping tool ecosystems reusable across runtimes. Note it ships in beta, API may still shift. (LangChain)
    • If you go with Bedrock Managed Agents for residency, confirm which orchestration features are OpenAI-hosted vs. customer-controlled before committing regulated workloads. (runtimewire)
Source shelf

Quick Links