AI digest: 2026-11-09

Highlights

• Claude Code 2.1.319 runs MCP tasks in the background, adds permissions.denyInsteadOfPrompting for unattended runs and /claude-api latency-optimize, and fixes several permission bypasses. Upgrade if you use auto or accept-edits mode.
• Anthropic's "The components of agent memory" gives measured advice. Tell strong models what memory is for rather than rules about what to save, keep the startup context small, and run a "dreaming" pass that corrects notes against transcripts.
• sandbox-mcp gives agents a throwaway Windows 11 machine to deploy, drive, and test apps in (Windows Pro or Enterprise only).

New sources

Agents and tooling

• Claude Code release 2.1.319: MCP tasks in the background, deny-instead-of-prompt, latency-optimize, security fixes.
• sandbox-mcp: CLI and MCP server for disposable Windows Sandbox test machines.
• GitHub Copilot weekly releases: November 2: CLI /settings dashboard, guided tool approval, Copilot in JetBrains Air.

Frameworks and libraries

• Claude Agent SDK (TypeScript) v0.3.319: MCP task support in SDK sessions.
• Google ADK (Python) v2.14.0: stable environments and A2A, multi-run evals, semantic tool-argument matching.
• Ollama v0.40.15: MiMo V2.6 coding model, AutoTrust decision models, Kimi Code CLI in ollama launch.

Models

• Gemini Embedding 2 Preview deprecated: routed to gemini-embedding-2, same embedding space, no re-indexing.

Research

• The components of agent memory: Anthropic's layered memory model, dreaming, and experiment results.

Topic pages updated

• agents/memory: new "Layers of memory" section and measured guidance; open question on what to remember largely answered (gap filled).
• agents/agent-patterns: memory layering and partitioned writes.
• practices/prompting: keep always-loaded instructions short; purpose over rules.
• practices/evals: multi-run evals and semantic argument matching (ADK 2.14).
• tooling/coding-agents: Claude Code 2.1.319 auto-mode and permission changes, Copilot updates.
• tooling/mcp: MCP tasks client support, sandbox-mcp.
• practices/security: Claude Code 2.1.319 bypass fixes, disposable test machines.
• frameworks/landscape: ADK 2.14.0, TS SDK 0.3.319, Pydantic AI 2.66.0 rows.
• models/open-weight: Ollama 0.40.15 models.
• models/gemini: Gemini Embedding 2 Preview deprecation.
• fundamentals/embeddings: when switching models needs no re-embedding.