Bloomberg reports Anthropic finished diligence then dropped a roughly $6B deal for Decart AI, whose chip-efficiency stack was meant to squeeze more training and inference from existing compute ahead of Anthropic’s IPO. The sides may still pursue a commercial partnership instead of an acquisition.
In an Oct 22 guide, Anthropic's Applied AI team says a single agent loop with on demand skills and MCP tools often beat router plus specialist designs on quality, cost and latency across enterprise deployments, so keep a subagent only for measured context bloat, parallel throwaway work, tool choice trouble or direct user contact. The Apache 2.0 single agent loop kit walks you through it: build a 10 example eval, fold one subagent at a time, and keep any whose replacement scores worse…
Announced Oct 22, new sessions in most Claude apps will stop asking you to approve each step; a classifier lets safe actions through and blocks risky ones while Claude looks for another route, and you can switch a session back to manual anytime. It's the same default Claude Code got in August, and Enterprise admins get an email before it reaches their orgs…
Shown on HN Oct 23, the MIT licensed, macOS only server controls a running Photoshop over AppleScript with about 20 tools for layers, adjustment layer grading, Select Subject, masks, text and export, and every edit is one undo step that rolls back automatically if it fails. Install from source for now ( ./scripts/install.sh client claude code ), since the 1.2.0 DMGs are unnotarized (build notes)…
Released Oct 22, it adds TypesafeClassifier and ClassifierNode for routing workflows without an LLM call, live audio with barge in, and experimental MongoDB session and keyword match memory services that need no vector index. Heads up: telemetry now defaults to OTel GenAI v2 and drops the call llm span, so set ADK TELEMETRY SCHEMA VERSION OPT IN=otel semconv 1 36 if your dashboards depend on the old schema…
Released Oct 23, @modelcontextprotocol/server/ext/skills lets a server load a skills folder with loadSkills(folder) and serve it via skills/list , skills/get and skill:// resources, while the matching SkillsClient verifies each file's size, digest and frontmatter. Heads up: x mcp header on number typed tool params is now rejected per the 2026 07 28 spec, so switch those to integers…
Spotted something we missed? Start a thread.
