Weekly signal

This week (June 22–30, 2026) accelerated a practical shift from "LLMs as answer engines" to "agents that work with people in context." Vendors pushed agent capabilities into collaboration and control surfaces (Slack, browser/desktop) while regulators and labs tightened who may run the most capable agents. The upshot: teams can build more autonomous assistants, but must pair them with containment, explicit human-in-the-loop patterns, and contingency plans for access restrictions.

What changed

1. Anthropic launched "Claude Tag," a persistent AI teammate that you can @mention inside Slack to surface context, summarize threads, and act on approved tools/data—available in research preview for Team and Enterprise customers. It treats Slack channels as shared agent workspaces rather than one-on-one chats. (Announced June 23, 2026).

2. Google made its Computer Use tool a native, public‑preview capability inside gemini-3.5-flash, letting a single Gemini agent perceive screens and take actions across browser, mobile, and desktop without stitching to a separate model. This is a developer-facing, agentic building block (announced June 24, 2026).

3. OpenAI previewed GPT‑5.6 (Sol/Terra/Luna) with a new ultra/subagent mode and a limited, government‑coordinated preview. GPT‑5.6 surfaces stronger multi‑step and multi‑agent reasoning intended for long‑horizon workflows; access is initially gated. (Preview published June 26, 2026).

4. The U.S. government partially restored Anthropic's Mythos 5 for ~100 vetted U.S. institutions after a June 12 export-control suspension; Fable 5 remains suspended. That precedent makes access to frontier agent capabilities a policy variable enterprises must manage. (Reported June 26–27, 2026).

What to do with it

• Treat agents as platform features, not just endpoints. Prototype shared/team agents (Slack Tag, persistent contexts) for workflows where institutional memory and asynchronous handoffs matter; start with read-only connectors and explicit scoping.
• Bake containment into design. Use VM/sandbox egress rules, principle-of-least-privilege connectors, and explicit approval patterns to avoid approval fatigue and persistent memory poisoning. Follow Anthropic’s containment patterns (ephemeral containers, human‑in‑the‑loop sandboxes, sealed VMs).
• Design multi-model fallbacks and vendor-continuity plans. Government gating (Mythos 5) shows access can change overnight—architect for model portability and graceful degradation.
• For production agents that act (computer use): require audit trails, action confirmations, and test prompt‑injection defenses in staging before any write or payments-capable flows.