OpenAI moved its internal agent-orchestration stack into public beta this month, and it's worth pausing on what that actually hands developers versus what it doesn't. The new managed Agents API handles orchestration, long-running sessions, and context management for autonomous agents, with sandbox compute sourced from OpenAI itself or from partners like Vercel and DigitalOcean. An OpenAI engineer also confirmed that the scaled-agent infrastructure powering ChatGPT's work product is now exposed as a public API, with setup times under a minute for spinning up agents on demand. That's a meaningful shift in where the orchestration layer sits. Until now, most teams building multi-agent systems assembled their own session management, retry logic, and sandboxing — often duct-taping together LangGraph, custom queues, and whatever compute they could provision. OpenAI is now offering to own that layer directly, the same way cloud providers absorbed load balancing and autoscaling a decade ago. If it works reliably at scale, it collapses a lot of custom infrastructure that startups have been building and rebuilding. The catch is the one every orchestration announcement this year eventually runs into: token delegation across agent chains. When Agent A calls Agent B, which calls Agent C, C needs proof that A actually authorized the original chain — and that's a governance question, not a compute question, and it's not what a managed sandbox API solves. Coverage of this same period notes the broader industry is converging on the Agent-to-Agent protocol for that inter-agent communication, with more than 150 organizations now participating in its governance, but the delegation-chain problem is explicitly called out as the piece nobody has closed yet. So the practical read: OpenAI's beta makes it dramatically easier to stand up production-style agents without building your own session and sandbox plumbing. It does nothing, by itself, to prove that an agent three hops deep in a workflow was actually authorized to be there. Teams adopting the managed API still need an external attestation or identity layer if they want an audit trail that survives a dispute — the orchestration convenience and the accountability question are separate problems, and vendors bundling agent infrastructure this month are mostly solving the first one. Worth watching whether OpenAI or a third party addresses delegation proof as a first-class feature in the next few beta iterations, or whether it stays a gap that A2A, MCP, and identity-layer projects have to fill from outside.