The Intake — Sunday, September 21, 2026

If you've been assembling session orchestration by hand — managing context windows, writing handoff logic between subagents, dispatching parallel tool calls — OpenAI's Agents API entered public beta.

The API opened September 10. It handles long-running sessions with automatic context compaction as conversations approach token limits. It supports parallel tool execution and multi-agent subagent coordination. Built-in integrations include MCP server connections and web search. Hosted sandbox options are available through partners Cloudflare, Oracle, Vercel, E2B, and Modal. OpenAI says there is no separate harness fee beyond token and tool costs.

Three early adopters reported outcomes at launch. SafetyKit reported a 60% cost reduction. Hypha reported an 86% reduction in failure rates. Cirridae reported evaluation scores rising from 0.71 to 0.85. All three figures are OpenAI-reported at launch; no independent verification has been published. If you're building multi-step agent workflows on custom session infrastructure, this is the managed path that became publicly accessible this week.

The "Pace the Frontier" coordination — Dario Amodei's essay, the public agreements from Altman, Musk, and Hassabis — became a federal antitrust complaint this week.

Four paid subscribers filed a Sherman Act §1 class action in the Northern District of California on September 18. The case is Buist v. Anthropic PBC; defendants are Anthropic, OpenAI, SpaceXAI, and Google. The complaint's central evidence is a September 12 coordination sequence. Amodei published "We Must Pace the Frontier" on September 12. The essay proposes embedded third-party evaluators with employee-level lab access and contractual publication rights for their findings. All three other lab CEOs — Altman, Musk, and Hassabis — publicly agreed that same day. OpenAI policy chief Chris Lehane separately confirmed an Anthropic-OpenAI-Google DeepMind working group. It formed in July 2026 and was still meeting in mid-September. The complaint alleges a horizontal agreement to restrain the pace of competitive product improvement. The proposed class is subscribers to all four labs' AI services.

If you subscribe to Claude, ChatGPT, Grok, or Gemini, you're in the proposed class. The complaint's theory is that coordinated pacing reduced the competitive product advancement those subscriptions were sold to deliver.

If you've been building in Kotlin or for Android and deferring agent functionality because Google's ADK wasn't at production parity in Kotlin, the 1.0 release is that milestone.

Google released ADK for Kotlin 1.0 on September 9. The release reaches feature parity with the Python and Java ADK versions. For Android, it adds on-device AI. The implementation runs through LiteRT-LM and ML Kit. Firebase AI Logic handles hybrid cloud-plus-device inference for Android deployments. The release supports hierarchical multi-agent task delegation with parent-child agent architectures. Compile-time tool schema generation via KSP annotations avoids the runtime reflection penalty. Automatic context compaction and history summarization are built in. Human-in-the-loop confirmation workflows are supported.

If you're building Kotlin or Android apps that need agent functionality, Google says the 1.0 release is the point where the ADK reached production-ready feature parity. On-device inference and hybrid cloud-plus-device deployment are now part of that baseline.

---