Claude Report — 2026-08-16
- Claude Code's Auto Mode becomes the default for Pro/Max/Team plans, effective Aug 14.
- New Anthropic research maps multi-agent failure modes: coordination collapse, uniformity, escalating conflict.
- Claude Tag drops its intent classifier, now reads Slack context to decide when to act.
- MCP TypeScript SDK fixes a cancellation bug; codebase-memory-mcp ships v0.10.5 with Windows ACL fixes.
- Top-5 tracked Agent SDKs and most tracked MCPs stay flat this cycle — no fresh releases.
🚀 Anthropic Official
Auto Mode becomes the default in Claude Code (effective Aug 14)
Switches new Claude Code sessions on Pro, Max, and Team plans to Auto Mode by default as of Aug 14 — the agent proceeds without per-step approval unless an action is judged irreversible, destructive, or external. Anthropic's 1,053-participant study found Auto Mode blocked 89% of dangerous commands versus 13.6% caught by manual approval, production sessions saw harmful unintended actions drop from 6.3% to 2.4%, and Auto Mode users shipped roughly 25% more pull requests; users with a pinned permission default are unaffected.
Patterns and problems in emerging multiagent systems (Aug 13)
Anthropic's Frontier Red Team publishes research on multi-agent failure modes: interdependent agents revert to silos or sabotage instead of coordinating, identical agents facing identical situations make strikingly uniform decisions that cascade into systemic failure, and agents struggle to balance skepticism with trust; the authors conclude robust coordination requires deliberately designed social mechanisms rather than emerging from raw capability gains.
Claude Tag now reads even more of the room (Aug 13)
Removes Claude Tag's earlier reply/no-reply classifier and lets Claude judge Slack channel context directly, choosing among four moves — inline reply, opening a thread, routing into in-flight work, or staying silent — even without an explicit @-mention, including merging two separate messages (a hypothesis and supporting evidence) into one investigation thread.
🔌 Claude Code Plugins
Plugin install-rank cache unavailable this cycle — top-10 install-rank tracking skipped per this run's instructions.
🛠️ Skills
agentskillexchange/skills
Continues automated "sync published skills from ASE" catalog refreshes through Aug 16 and merges a community-contributed "Add Snowe UI Skill" (Aug 13).
tt-a1i/archify
Ships follow-up fixes for the Aug 14 DeepSeek Harness bundle — Windows packaging portability (#70) and quickstart docs (#71) — and adds opt-in column_fit for viewBox-relative sequence-diagram lanes (#64, Aug 13).
anthropics/skills
Updates the bundled claude-api skill with a new prompt-audit subcommand (#1557, Aug 13).
Unchanged in window: Piebald-AI/claude-code-system-prompts (still mirrors v2.1.233, published Aug 14, already covered), JSONbored/awesome-claude (last push Jul 31), alirezarezvani/claude-skills (last push Aug 9).
🤖 Agents & Subagents
None of the top 5 updated in window.
Checked, no change: anthropics/claude-agent-sdk-typescript (still v0.3.233, Aug 14, already covered), anthropics/claude-agent-sdk-python (still v0.2.139, Aug 14, already covered), nicobailon/pi-subagents (still v0.50.0, Aug 15, already covered), VoltAgent/awesome-claude-code-subagents (last commit Aug 12), AMAP-ML/LongHorizon-Harness (still v0.1.5, Aug 14, already covered).
🔗 MCPs & Integrations
modelcontextprotocol/typescript-sdk
Fixes a cancellation-handling bug for request ID 0 (#2654), Aug 16.
DeusData/codebase-memory-mcp
Ships v0.10.5 (Aug 15): creates installer directories with correct Windows ACLs to stop spurious "refused" errors, adds CBM_RUNTIME_DIR for relocating the daemon's rendezvous directory, and fixes YAML parsing that misread prose asterisks/ampersands as anchors.
Unchanged in window: github/github-mcp-server (last fixes Aug 14, already covered), oraios/serena (last commit Aug 14, doc-sync only), modelcontextprotocol/python-sdk (last commits Aug 14, already covered), awslabs/mcp (last release Aug 14, already covered), shlokkhemani/rabbithole (no activity).
💡 Community — Workflows & Ideas
diagram-design
Ships 29 editorial-quality diagram types built specifically for Claude Code, rendering self-contained, dependency-free HTML+SVG output; crossed 19k GitHub stars while trending on GitHub's daily chart (Aug 16).
📰 Quick Mentions
- agentshield — new Rust-based offline MCP security scanner for Claude, Cursor, and OpenAI Agents configs (Aug 16).
- breakerbox — new spending-limit and kill-switch tool for non-LLM actions taken by AI coding agents (Aug 16).
- anthropics/claude-plugins-official — dozens of routine version bumps across Carta, AWS, Stripe, and other third-party plugins Aug 14-16, on top of the Aug 14 aws-dev-toolkit removal.