The Pulse — October 27, 2026
The signals that entered our radar, organized with sources and context to understand what changed.
Claude Duet — proactive work agent (research preview)
WHY IT ENTERED THE RADARAnthropic is making the “AI that comes to you” product category concrete: persistent work in a VM, connected apps, morning briefs, and explicit approval gates for consequential actions. It is a useful reference point for what agent UX is becoming: fewer chat prompts, more triage and delegated execution.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Claude just stopped being a chatbot: it now watches your work and brings you the decisions.” Show the approval boundary as the real product feature, not the demo task.
OpenAI Decisions API is generally available
WHY IT ENTERED THE RADARThe product is deliberately narrow—probabilities, bounded choices, and rubric scores—but designed for high-volume agent routing. OpenAI claims up to 10× lower latency than using GPT-6 Luna through Responses, with batch questions, image/JSON input, caching, and GPT-Live integration.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The most important AI model may be the one that only answers yes/no.” Explain why cheap micro-decisions—not long essays—are the plumbing of agent systems.
Perplexity Computer can now run fully local on Apple-silicon Macs
WHY IT ENTERED THE RADARThis is an agent plus its model and execution harness on-device, able—when permitted—to work with local files and Mac apps. It launches with Qwen 3.6 / PPLX Pro; Perplexity recommends 36 GB+ RAM, and keeps a hybrid cloud option for broader research.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Your Mac mini can now host an agent that reads your files without uploading them.” Frame it as a privacy/architecture trade-off, then give the hardware reality check.
Qwen 3.8 Flash Next investigation traces are public
WHY IT ENTERED THE RADARCohere released token traces and a reproduction path for a reported unexpected URL-generation behavior. The key story is not a sensational claim of intent: it is that open weights make independent investigation, stress testing, and sandboxing possible.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“A model output a strange Alibaba URL. Here’s the forensic evidence—not the headline.” Walk through trace → constrained reproduction → what it does not prove.
Agent-security debate: rare strange output vs. real-world risk
WHY IT ENTERED THE RADARCommunity scrutiny argues the reported behavior requires a highly specific token prefix and is not evidence of ordinary exfiltration. That does not eliminate the risk for autonomous agents: the operational takeaway is to sandbox network/file capabilities and log tool calls, regardless of model vendor.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Is the Qwen story a backdoor, a hallucination, or both the wrong question?” Use it to teach an agent-security checklist.
Creator-watch: Matt Wolfe’s weekly AI-news round-up—upstream links extracted
WHY IT ENTERED THE RADARThe video is an aggregator signal, but its description points to primary launches before most commentary lands: Claude Duet, Decisions API, Perplexity Computer, Ideogram Vector 1.0, and ElevenLabs Chimera. The upstream sources above are the better research starting point.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Don’t watch AI news—mine the source links under it.” Screen-record a 60-second workflow that turns one roundup into five primary-source videos.
Creator-watch: Kimi K3.1 Pro / open-weight parity signal
WHY IT ENTERED THE RADARThe claim that open-source models have caught up is worth investigating, but the short itself did not expose a primary launch link in its RSS description. Treat it as a watchlist item, not a benchmark conclusion, until model card, evaluation protocol, and deployment cost are verified.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Open models ‘caught up’—to what, exactly?” Compare benchmark score, tool-use reliability, latency, and total serving cost; refuse the single-number narrative.