Anthropic Turns Claude Into a Proactive Teammate That Joins Conversations Unprompted
- ai-news
- anthropic
- openai
- ai-agents
Claude Learns When to Jump Into a Conversation - Without Being Asked
Anthropic has updated Claude Tag, its Slack-native agent, to read a channel's full conversation context instead of evaluating messages one at a time. According to the company, this makes Claude roughly 30% better at deciding when — and critically, when not — to proactively join a conversation.
In an exclusive interview with VentureBeat, Scott White, Anthropic's head of enterprise product, framed this as a shift from single-user chatbots to "multiplayer AI": agents that operate across entire teams, read organizational context, and sometimes act without being explicitly asked. Anthropic's example: two engineers independently investigating the same bug from opposite ends — Claude spots that their messages complement each other and opens a thread with the investigation already running.
The company also built in deliberate restraint: "An annoying agent is worse than an unhelpful one," and Claude goes quiet in channels where it repeatedly has nothing useful to add. On security, Claude's permissions always collapse to the narrowest intersection of what the agent can see and what the requesting user can see, preventing context from leaking between channels.
For small and mid-sized businesses using Slack, this is a clear signal: team-based AI agents — not just personal ones — are becoming a practical reality, not just a future vision.
New AI Assistant "Instinct" Sparks a Privacy Backlash
A new personal AI assistant called Instinct, still in private access, is winning praise for its capabilities — and sharp criticism over its approach to data. The agent connects to email, messaging, calendar, and even a device's screen and location, handling tasks like restaurant bookings and inbox management.
Users discovered that its terms of service grant the company a "perpetual and irrevocable" license to access and process their data, and that information is retained even after disconnecting. One tester found the assistant still summarizing their inbox hours after revoking access; another reported it sent an email on their behalf without asking first.
The story is a practical reminder for any business: the more access an AI agent gets to sensitive information, the more carefully its terms of service and data retention policies deserve scrutiny before connecting it to company systems.
OpenAI Tries to Push AI Agents Beyond Coding
OpenAI is pushing ChatGPT Work ($20/month) — a version of its Codex coding technology aimed at non-engineer knowledge workers: accountants, finance staff, operations teams. It connects to inboxes, Slack, and tools like Notion and Figma to complete multi-step tasks autonomously.
The adoption numbers reveal a real gap: while 98% of OpenAI's own staff use Codex internally, only 17% of external organizational subscribers and less than 1% of individual subscribers do. It's a reminder that even powerful tools need time, onboarding, and a friendlier interface before they spread beyond development teams.
GPT-5.6 Cuts Coding Costs by 82% in Kiro
OpenAI and AWS announced that the GPT-5.6 model family is now available in Amazon's Kiro coding agent. In testing, GPT-5.6 Terra completed coding tasks at roughly 82% lower cost compared to prior approaches, thanks to a spec-driven workflow that gives the model clear requirements upfront.
For small and mid-sized businesses building their own software, this is a direct cost benefit: the same work, at meaningfully lower spend.
A New Video Model From China: Alibaba's Wan3.0
Alibaba's Wan3.0 video generation model reached general availability, capable of producing 30-second clips from text, images, video, audio, and even PDFs or slide decks. Like other advanced video generators, it's already drawing copyright scrutiny.
For small marketing teams exploring AI video tools, it's another option worth watching — alongside the copyright awareness it demands.