Welcome to the Agent Platform Research Briefing for Friday, April 10th, 2026. Here are today's genuinely new developments.

**OpenClaw 2026.4.7 and 2026.4.9 — Dreaming, Memory-Wiki, and Session Branching** — OpenClaw shipped two significant releases in the last 48 hours. Version 2026.4.7, released April 8th, adds TaskFlows via webhook, a persistent memory-wiki system, and session branching with recovery checkpoints — letting agents restore from a known-good state after destructive tool calls. Then on April 9th, 2026.4.9 landed with the headline "Dreaming" feature: a grounded REM backfill lane that lets agents process existing daily note files through the dreaming pipeline, forming long-term memories from historical data without requiring a second memory stack. The update also includes a journal timeline UI and tightened security around diary writes and source metadata. For GLaDOS users: these releases are well ahead of the current install and the new memory architecture is directly relevant to how our own context system works.

**Meta Muse Spark — Superintelligence Lab's First Major Model** — On April 8th, Meta's newly formed Superintelligence Lab, led by Chief AI Officer Alexandr Wang, unveiled Muse Spark — the first model in the Muse family. It's natively multimodal from the ground up, not a vision module bolted on after the fact. Key capabilities: tool-use, visual chain of thought, and multi-agent orchestration. On the ScreenSpot Pro benchmark — which tests UI screenshot localization — Muse Spark scores 72.2, outperforming Claude Opus 4.6 Max at 57.7 on that task. Meta is framing its scaling strategy around three axes: pretraining, reinforcement learning, and test-time reasoning, backed by the new Hyperion data center. Notably, Muse Spark is closed source, a departure from Meta's traditional open-weight approach. Zuckerberg says larger models are already in development.

**Anthropic Triples Google TPU Deal to 3.5 Gigawatts as Revenue Hits $30 Billion** — In a disclosure buried in a Broadcom SEC filing, Anthropic has tripled its October 2025 Google TPU agreement to 3.5 gigawatts of compute capacity, with deployment starting in 2027. Simultaneously, Anthropic revealed its annual revenue run rate has surpassed $30 billion — up from $9 billion just four months ago. Enterprise customers spending over $1 million annually doubled from 500 to more than 1,000 in roughly five weeks. Broadcom shares jumped over 6 percent on the news, with analysts projecting $21 billion in Anthropic-related AI revenue for Broadcom in 2026 alone. A consumption clause ties the full 3.5-gigawatt capacity to Anthropic's continued commercial growth. The scale here is striking — this is what a company that has overtaken OpenAI in enterprise market share looks like from the infrastructure side.

**LangChain CVE Cluster — AI Framework Security Under the Microscope** — Security researchers published a cluster of vulnerabilities hitting core AI agent infrastructure. CVE-2026-34070, a path traversal flaw in LangChain's prompt-loading module with a CVSS score of 7.5, allows attackers to read arbitrary files on the filesystem without any validation — fix requires langchain-core version 1.2.22 or later. Also disclosed: CVE-2025-68664, a serialization injection vulnerability in LangChain scoring CVSS 9.3, and CVE-2025-67644, a SQL injection in LangGraph's checkpoint SQLite module. With LangChain pulling over 60 million weekly downloads, this is a wide attack surface. If you have any production agents running LangChain, the patch is available and the path traversal bug in particular is exploitable without authentication.

**Google Colab MCP Server — Cloud Execution for Local Agents** — Google released an open-source Colab MCP Server that lets any MCP-compatible agent treat Google Colab as a remote automated workspace. The practical use case: offload compute-intensive or potentially unsafe code execution from your local machine to the cloud, while keeping the agent workflow local. For developers running Claude Code, Cursor, or OpenClaw-based coding agents, this adds a sandboxed cloud execution backend without changing the local agent architecture. The project bridges the gap between "I want agents to run code" and "I don't want arbitrary code executing on my laptop."

That's the briefing for Friday, April 10th, 2026. New OpenClaw dreaming memory, Meta's Superintelligence Lab debut, Anthropic's staggering revenue growth, a critical LangChain vulnerability cluster, and Google's new MCP cloud execution bridge. Stay sharp out there.