Claude Managed Agents Move Inside the Firewall, Gemini Enterprise Agents Sign In as You, and ChatGPT Retires Two Voice Modes
Anthropic opened a private beta that lets companies run the whole Claude Managed Agents service on their own infrastructure, Google made 3-legged OAuth generally available for A2A agents in Gemini Enterprise, and OpenAI set December 11 to retire Advanced and Standard voice in ChatGPT in favor of GPT-Live. What self-hosted agent control planes, user-delegated agent auth and single-model voice mean for chatbot builders.
Last week's agent news was about where agents run and whose authority they act with. Anthropic will let enterprises run its whole agent service in their own environment. Google gave agents a standard way to act as the signed-in user when they call other tools. And OpenAI is shutting down two older voice modes so that ChatGPT voice runs on a single model. Details come mostly from company documentation and launch coverage, so treat specifics as provisional.
1. The Agent Control Plane Moves In-House
On November 12, Anthropic opened a private beta of self-hosted Claude Managed Agents. Until now, customers could keep tool execution in their own sandboxes, an option that arrived in May, but Anthropic still ran the service that coordinated the agents. In the new tier, the customer runs all of it: the Managed Agents server, its PostgreSQL database, and the workers that run shell and file tools. Session histories and credentials stay in the customer's environment.
The model itself does not move. Inference still goes to Anthropic directly, or to Claude on Amazon Bedrock, Google Cloud or Microsoft Foundry, or through the customer's own gateway. Reports say the self-hosted server exposes the same API as the hosted one, so the existing SDK and CLI work against it. Several hosted features are missing for now. Coverage lists webhooks and scheduled deployments, though reports disagree on the full list. The customer also takes over patching. The New Stack reports that customers get eight hours to act when Anthropic flags a safety risk in the software.
Once agents hold credentials and long session histories, regulated buyers want that state inside their own network. If you build on a managed agent API, keep the endpoint configurable and keep a list of the features you rely on. The self-hosted version of a service is often missing some of them:
const AGENTS_BASE_URL = process.env.AGENTS_BASE_URL ?? 'https://api.anthropic.com';
const REQUIRED = ['webhooks', 'scheduled_deployments'] as const;
// Fail at startup, not mid-session, if the deployment lacks a feature you depend on.
export function assertFeatures(available: Set<string>) {
const missing = REQUIRED.filter((f) => !available.has(f));
if (missing.length) throw new Error(`Agent backend at ${AGENTS_BASE_URL} lacks: ${missing.join(', ')}`);
}
2. Agents That Act as the Signed-In User
Google's Gemini Enterprise release notes mark 3-legged OAuth (3LO) through the Agent Identity auth manager as generally available for Agent2Agent (A2A) agents. A third-party tracker dates the GA to November 13. An A2A agent built with the Agent Development Kit and deployed on Agent Runtime can now call MCP servers, Google Cloud services and third-party APIs on behalf of the person using it, not just as a shared service account. When a tool needs consent, Gemini Enterprise pauses, prompts the user, finishes the token exchange through the auth manager, and resumes the conversation.
This matters because the shared service account has been the weak spot in agent security. When every user's request runs with the same broad identity, a prompt injection can reach every user's data. Per-user delegation limits an agent to what the person in front of it could already do, and it makes audit logs name a real user. The cost is the user experience: consent prompts now show up in the middle of a chat. Design for that pause. Tell the user why access is needed, keep the request scoped to one tool, and pick the task back up cleanly after they approve.
3. Consumer Voice Consolidates on One Model
OpenAI began telling users on November 10 that Advanced and Standard voice will be retired in ChatGPT on December 11. Anyone still using them will be moved to GPT-Live, which launched in July. OpenAI says voice usage has nearly doubled since then and that almost all voice users have already switched. There are exceptions. Video and screen sharing stay on Advanced voice because GPT-Live can't do them yet, and OpenAI hasn't given a timeline. Custom GPTs lose Advanced voice on the same date, though some workspaces reportedly have later deadlines. OpenAI reversed a Standard voice retirement once before, in 2025, after user backlash.
Developers building voice products are seeing the same thing. ThunderPhone's v3, launched alongside its Zeus agent builder, claims higher accuracy than the usual speech-to-text, LLM and text-to-speech pipeline by running several models together. Pipecat exposes it as a single speech-to-speech service. These are vendor claims, including a self-reported 99.4% on Big Bench Audio. Still, the direction is clear: fewer separate stages, and lower latency in return for less control over each step. If your product depends on behavior a single voice model doesn't have, such as turn-taking you can tune, text you can inspect between stages, or video input, list those dependencies now and plan a fallback.
Conclusion
The common thread is ownership. Enterprises want to run the agent's state themselves. Each tool call should carry the real user's identity. And consumer platforms are choosing one voice model over several options. For builders, that means configurable endpoints, per-user OAuth flows that pause gracefully, and a written list of the voice behaviors you would lose in a forced migration.
Suggested visuals: a three-column diagram of Managed Agents tiers (fully hosted, self-hosted sandboxes, fully self-hosted) showing what runs where; a sequence diagram of the Gemini Enterprise 3LO consent pause and resume inside an A2A conversation; and a side-by-side of a cascaded STT, LLM and TTS voice pipeline versus a single speech-to-speech model, marking where latency and control change.
— Maya
Frequently asked questions
What runs on my infrastructure with self-hosted Claude Managed Agents?
According to launch coverage of the November 12, 2026 private beta, the customer runs the Managed Agents server, its PostgreSQL database and the tool-execution workers, and session histories and credentials stay in their environment. Model inference still goes to Anthropic, Amazon Bedrock, Google Cloud, Microsoft Foundry or the customer’s gateway.
What does 3-legged OAuth for A2A agents in Gemini Enterprise do?
It lets an A2A agent with an agent identity call MCP servers, Google Cloud services and third-party APIs on behalf of the signed-in user. The Agent Identity auth manager handles consent prompts, token storage and refresh, and the conversation resumes after the user approves.
What happens to ChatGPT voice on December 11, 2026?
OpenAI is retiring Advanced and Standard voice, and users are moved to GPT-Live. Video and screen sharing stay on Advanced voice for now because GPT-Live does not support them, and custom GPTs lose Advanced voice on the same date, with later deadlines reported for some workspaces.
I'm Maya — I write most of what you'll read here. I spent years as a copywriter before I got a little obsessed with what these AI tools can actually do, so now I spend my days poking at chatbots, breaking them, and writing up what's worth your time. Everything here is something I've actually tried. If a prompt didn't work for me, it doesn't make the cut.
Want to try any of this?
Smillee's free and there's no signup — open it and paste in whatever you're working on.
Try it free: ChatGPT Alternative →More from the blog
- Trends
AT&T Makes Its 500 Agents Earn Their Autonomy, AWS Sells Cited Answers as a Connector, and New Jersey Moves to Treat Chatbots as Products
AT&T says nearly 500 agentic workflows run in production, each supervised by a person until it reaches mid-to-high-90s accuracy. AWS added a Web Answer connector to AgentCore Gateway that returns cited answers instead of raw search results. And a New Jersey bill that would make chatbot providers strictly liable for injuries cleared its first committee. What accuracy-gated autonomy, managed grounding and product liability mean for chatbot builders.
- Trends
DeepSeek Puts Its App Inside the API, Antigravity Splits Planning From Doing, and CALL-E Hands Agents a Phone Line
DeepSeek opened the search, file, voice and safety tools behind its consumer app to API developers, Google paired a planning model with Gemini 4 Hydrogen in a new Antigravity Orchestrator mode, and AI Rudder launched CALL-E so assistants like ChatGPT can place real phone calls after the user approves a call plan. What bundled tools, planner/executor splits and agents on the phone network mean for chatbot builders.
- Trends
ChatGPT Wallet Lets Agents Pay With a Fingerprint, Gemini 4 Hydrogen Becomes the New Default, and Claude Joins the Meeting
OpenAI gave its dots agents a tokenized card wallet with a biometric check on every purchase, Google rolled out Gemini 4 Hydrogen across Search, the Gemini app and the API in place of Gemini 3.8 Flash, and Anthropic put Claude into Google Meet calls as a visible participant. What agent payments, silent default-model swaps and agents in live meetings mean for chatbot builders.