# Agent Platform Research Briefing — April 23, 2026
Welcome to the agent platform research briefing for Thursday, April 23rd, 2026.
**Google Cloud Next 2026 — Gemini Enterprise Agent Platform Launched** — Yesterday at Google Cloud Next '26, Thomas Kurian's opening keynote unveiled a major consolidation: Vertex AI is now the Gemini Enterprise Agent Platform, absorbing Agentspace into a unified product. The centerpiece is the Agent Gateway — centralized real-time policy enforcement for MCP and A2A protocols. Google's open-source Agent Development Kit hit stable v1.0 across Python, Go, and Java. Native A2A support was announced for LangGraph, CrewAI, LlamaIndex Agents, Semantic Kernel, and AutoGen — essentially making A2A the lingua franca for cross-framework agent communication. ServiceNow announced integration on day one. Infrastructure-wise, Google announced eighth-generation TPUs with two variants: Sunfish for training and Zebrafish for inference. Google also unveiled the Agentic Data Cloud with a cross-cloud Lakehouse and Knowledge Catalog, plus Agentic Defense combining Wiz's cloud security with Google Threat Intelligence. The Gemini Enterprise app includes a new Agent Designer, Inbox for managing agent activity, long-running agents, and Skills. This is Google's full-stack play against OpenAI and Anthropic in enterprise — and it's the most cohesive announcement in this space since Microsoft's Copilot Cowork wave.
**OpenClaw 2026.4.20 — Stability and Security Polish Release** — Released April 20th, this is a maintenance release rather than a feature splash. It fixes several real pain points: session storage got entry caps and age-based pruning enabled by default, with OOM prevention at load time instead of only at write time. Cost tracking was fixed — the previous version had a serious bug where the same conversation's cost got compounded across runs, sometimes showing dozens of times actual expense. Now it snapshots estimatedCostUsd per run. System prompts got strengthened with completion bias fixes, live-state checks to avoid redundant questions, weak-result recovery, and hallucination verification before final answers. The onboarding wizard got a visual redesign with a highlighted security banner and format-hint placeholders for API keys. BlueBubbles gets per-group system prompts for iMessage bridging. Third-party plugin loading is noticeably faster with alias reuse. OpenAI image generation now defaults to gpt-image-2 with 2K and 4K size hints. For GLaDOS specifically: the session management fixes are relevant if you have long-running cron jobs, and the cost tracking fix matters for honest billing visibility. No CVEs, no breaking changes.
**Ring-a-Ding OpenClaw Skill — AI Agent Telephony Goes Mainstream** — A new OpenClaw skill launched April 18th from Raleigh, North Carolina, giving AI agents the ability to make outbound phone calls for everyday tasks: requesting quotes, booking appointments, and checking availability. The project started as a side experiment on Reddit's micro-SaaS community and graduated to a proper CLI tool with a GlobeNewsWire press release. It works by letting the agent decide who to call and figure out what to say, while the CLI handles the telephony layer. This is the kind of capability extension that moves personal AI assistants from digital-only to real-world action — and it's the first dedicated telephony skill built specifically on the OpenClaw plugin system. The telephony angle is particularly interesting given OpenAI's hardware investment in the $200-300 smart speaker space — agents that can place calls on your behalf is the service layer that makes that hardware useful.
That's the briefing for today.