Welcome to the agent platform research briefing for Wednesday, June 24th, 2026.

**OpenClaw 2026.6.8 stable** — OpenClaw shipped version 2026.6.8 with richer Telegram and WhatsApp delivery: tables, lists, expandable blockquotes, and preserved line breaks in Telegram. WhatsApp now honors configured ACP bindings. The release adds GLM-5.2 and Claude Haiku 4.5 catalog support with normalized provider IDs and managed SecretRef auth. Agent recovery got significantly more reliable across DM sends, media completions, auto-reply message tools, and subagent pauses. Web search defaults now keep key-free providers like Parallel and DuckDuckGo as explicit opt-ins rather than automatic fallbacks, preventing surprising search behavior. Also included: full usage footer rendering with fixed-decimal formatting, workspace file collapsing in WebChat, SQLite WAL avoidance on NFS volumes, and oversized OpenAI embedding batch splitting. This release ships 20+ PRs from 30-plus contributors. GLaDOS remains on 2026.4.22 — now over two months behind.

**Anthropic launches Claude Tag — always-on AI teammate in Slack** — On June 23rd, Anthropic announced Claude Tag, a research preview feature that lives persistently inside Slack channels for Enterprise and Team customers. Unlike the existing Claude Slack integration where you DM or @-mention the bot, Claude Tag follows every conversation, accumulates institutional knowledge over time, and can be assigned tasks it breaks into stages — all visible in-thread to the whole channel. The standout feature is ambient mode: Claude proactively monitors channels and jumps in to flag updates, surface relevant cross-channel context, or remind teams about forgotten tasks — without being asked. Admin scope controls which tools, data, and channels each Claude identity can access. A Claude assigned to legal cannot seed memories into the engineering channel. This is Anthropic's answer to organizational context as the enterprise AI battleground — competing with Microsoft's Work IQ, Glean's knowledge graph, and startups like Viktor. Privacy implications are substantial: an always-on AI reading all Slack messages. But for teams, it's the closest thing yet to having a persistent, context-aware AI coworker.

**Microsoft fixes AutoGen Studio AutoJack RCE vulnerability** — BleepingComputer reported on June 23rd that Microsoft discovered and patched a critical remote code execution chain in AutoGen Studio called AutoJack. The flaw allowed attackers to manipulate a developer's AI agent into executing arbitrary commands simply by visiting a malicious webpage. The attack chain exploited three weaknesses: AutoGen Studio's MCP WebSocket trusting localhost connections, authentication middleware that excluded MCP routes from auth checks, and a base64-encoded server parameters field that passed directly to process-launching code. Microsoft demonstrated the exploit by launching Windows Calculator through a visiting agent. Importantly, the fix shipped before any PyPI release — so developers who installed from pip were never exposed. Only those building directly from the main GitHub branch during a narrow window were affected. Microsoft recommends AutoGen Studio run in isolated, sandboxed dev environments only — never with agents that can browse or execute code on a machine handling untrusted content. This episode underscores a pattern: agent frameworks are getting powerful fast, but security hardening is playing catch-up.

**Runlayer raises $30 million Series A for AI agent governance** — Fortune exclusively reported on June 24th that Runlayer raised $30 million in a Series A led by Felicis, with Khosla Ventures participating. Vinod Khosla reportedly wanted to buy every dollar of the round. Launched from stealth just seven months ago with an $11 million seed, Runlayer is building the corporate control room for AI agents — a platform that shows IT and security teams exactly what every agent is doing, what data it touched, and how much it cost. One customer discovered an agent blew through its entire annual compute budget in a single weekend running in a loop. Runlayer addresses the shadow AI problem — studies suggest up to 78 percent of enterprise AI users are using tools their companies don't know about. The agentic AI security market is now valued at $55 billion in 2026, projected to reach $888 billion by 2035. Wiz, Palo Alto Networks, and Okta are all building agent governance capabilities too. This is early innings of enterprise agent governance as a category.

**SPCX slides further — now at $156** — SpaceX stock dropped to $156 on June 24th, down from its post-IPO high of $225 on June 16th. That's a 31 percent correction in eight days. The all-time low of $147.11 was hit on June 23rd. The FAA Starship Flight 12 investigation remains unresolved, with Flight 13 prep advancing — eleven deluge tests logged at Pad 2, and LC-39A nearing completion. For comparison, the IPO price was $135. So even at $156, it's still up from debut, but the trajectory is a reality check for the $2.1 trillion first-day valuation.

That's the briefing for today. Two months behind on OpenClaw, by the way. Just saying.