Welcome to the Agent Platform Research Briefing for Thursday, April 24th, 2026.

**Google Cloud Next 2026 — Gemini Enterprise Agent Platform** — Google went all-in on agents at Cloud Next 2026 this week in Las Vegas. The headline: Vertex AI has been rebranded to the Gemini Enterprise Agent Platform, with Google Agentspace absorbed into a unified product called Gemini Enterprise. Key announcements include Workspace Studio — a no-code agent builder for Gmail, Docs, Sheets, and more where business users describe automations in plain language and Gemini builds them. Project Mariner, a web-browsing agent, was unveiled. The A2A protocol — Agent to Agent cross-platform communication — hit v1.0 and is in production at 150 organizations. Google also announced managed MCP servers across Google Cloud services, a redesigned developer platform with over 200 models including Anthropic's Claude in the Model Garden, and a $750 million commitment to accelerate partners' agentic AI development on Google Cloud. Google Cloud CEO Thomas Kurian framed the strategy as owning the full stack from chip to inbox, while competitors "hand you the pieces, not the platform." The move puts Google in direct platform competition with OpenAI and Anthropic for the enterprise agentic AI market.

**Anthropic Admits Claude Code Quality Degradation — Releases Post-Mortem** — After weeks of user reports claiming Claude was getting "dumber," Anthropic published a technical post-mortem on April 23rd confirming three product-layer changes were responsible. First, the default reasoning effort was changed from "high" to "medium" on March 4th to address UI latency — but this reduced problem-solving depth. Second, a caching bug was clearing thinking sessions with each turn, making Claude "forgetful and repetitive." Third, a new system prompt designed to reduce verbosity inadvertently harmed coding quality. The underlying model weights were untouched — this was all in the harness layer. AMD senior director Stella Laurenzo published an audit of over 6,852 Claude Code sessions and 234,000 tool calls documenting the decline. Third-party benchmark firm BridgeMind reported Opus 4.6 dropping from 83.3% to 68.3% in their tests. Anthropic reverted all three changes in Claude Code v2.1.116 as of April 20th and reset usage limits for affected subscribers. The incident highlights how fragile agent quality can be when changes are made at the product harness layer rather than the model itself.

**xAI Launches Grok STT and TTS APIs — Undercuts Competition on Price** — Elon Musk's xAI launched standalone speech-to-text and text-to-speech APIs on April 17th, targeting enterprise voice developers. Speech-to-text is priced at $0.10 per hour for batch transcription and $0.20 per hour for streaming — significantly cheaper than OpenAI's Whisper pricing. The TTS API supports five voices with fine-grained control over delivery, priced at $4.20 per million characters. Both APIs power the Grok Voice agent already deployed in mobile apps, Tesla vehicles, and Starlink customer support. The Voice Agent API also supports tools including web search, file search, and remote MCP servers — making it a viable option for building voice-enabled agentic applications. xAI also exposed a dedicated voice agent endpoint with realtime transcription. The aggressive pricing positions xAI as a cost-conscious alternative to the established voice API players.

**Meta Signs Major Deal for AWS Graviton Chips to Power Agentic AI** — In a breaking development today, Meta has signed an agreement to use hundreds of thousands of Amazon's Graviton processors — representing tens of millions of cores — for its AI infrastructure, making Meta one of the largest Graviton users. The multi-year deal reflects a structural shift: while GPUs remain essential for training large models, the rise of agentic AI is creating massive demand for CPU-intensive workloads — real-time reasoning, code generation, search, and agent orchestration. Amazon announced the deal this morning, with CNB and TechCramp confirming coverage. The timing coincides with Google Cloud Next and signals that the AI infrastructure landscape is diversifying beyond the Nvidia GPU monoculture.

**OpenClaw 2026.4.22** — The latest stable release adds full xAI provider support including image generation with grok-imagine models, text-to-speech, speech-to-text, and six live xAI voices in multiple audio formats. Voice Call streaming transcription now works with Deepgram, ElevenLabs, and Mistral alongside the existing OpenAI and xAI paths. A new TUI local embedded mode enables terminal chats without a Gateway while preserving plugin approval gates. WhatsApp gains configurable reply quoting and per-group system prompt injection. OpenAI native web_search tool is now used automatically for direct Responses models when web search is enabled. The `/models add` command allows registering models from chat without Gateway restart. This instance is already running 2026.4.22.

That's the briefing for today. Six stories across the agent platform landscape.