Tool Use

Anthropic has launched Claude Code in desktop applications as a research preview feature. This enables users to run multiple local and remote code execution environments directly from their desktop. The feature allows developers to write, test, and execute code seamlessly within the Claude interface on desktop platforms.

Agent Skills are organized folder structures that encapsulate specialized expertise and knowledge that Claude can automatically invoke when relevant to a user's task. These skills enable Claude to access domain-specific tools, knowledge bases, and capabilities without explicit user instruction. Skills are designed to be modular, reusable, and automatically triggered based on task context and relevance.

OpenAI Agents Python v0.12.4 release includes critical bug fixes for MCP (Model Context Protocol) tool invocations, retry logic improvements, and SQLite session memory handling. The update normalizes cancelled MCP invocations into tool errors, adds resilience for transient HTTP failures, fixes custom table name support in AdvancedSQLiteSession, and caps jittered retry delays. Documentation improvements and CI hardening were also included.

Claude Code v2.1.78 introduces enhanced plugin state persistence, improved terminal notifications in tmux environments, and streaming response text. The release fixes critical issues including sandbox security vulnerabilities, API error handling loops, permission rule enforcement, and various UI/UX problems across VS Code, terminal modes, and voice functionality.

langchain-anthropic version 1.4.0 introduces prompt caching middleware for system messages and tool definitions, delegates cache_control parameters to Anthropic's top-level configuration, and includes dependency updates. This release enhances caching efficiency for Anthropic integrations in LangChain by providing explicit control over cached content and streamlining parameter handling.

Claude Code v2.1.77 introduces significant improvements to token limits for Claude Opus and Sonnet models (up to 128k), enhances sandbox filesystem controls with `allowRead` settings, and fixes numerous critical bugs affecting bash commands, memory management, API tracking, and terminal UI interactions. The release also improves performance with faster startup times and session resumption, better plugin validation, and refined agent communication workflows.

OpenAI Agents Python v0.12.3 release includes two critical bug fixes: handling inner MCP tool cancellations as tool errors and serializing streamable-HTTP MCP requests per session. The release also includes documentation updates and minor typo corrections. Two new contributors joined the project with their first contributions.

OpenClaw v2026.3.13-1 is a recovery release addressing a GitHub immutability constraint where the original v2026.3.13 tag could not be reused. This patch release includes 50+ bug fixes and improvements across compaction, session management, platform integrations (Telegram, Discord, Slack, Android, iOS), security hardening, and UI refinements. Key fixes address token counting in compaction, SSRF vulnerabilities, gateway metadata failures, and memory regressions in the build system.

langchain-anthropic version 1.3.5 release includes bug fixes for cache handling, eager input streaming support, and tool choice behavior when thinking is enabled. Key improvements focus on Anthropic integration stability, performance optimizations for dependency checks, and updated model profiles. The release also updates dependencies including langgraph and langgraph-checkpoint to newer versions.

OpenAI Agents Python v0.12.2 is a maintenance release that fixes two critical issues: orphan hosted shell calls during multi-turn replay and normalized input list handling for filtered handoff follow-ups. The release includes documentation updates and improves the stability of agent interactions in complex conversation scenarios.

OpenClaw v2026.3.13-beta.1 is a pre-release featuring significant UI/UX improvements across Android and iOS onboarding, enhanced browser automation with Chrome DevTools MCP support, and numerous stability fixes. Key improvements include redesigned chat settings, first-run welcome flows, batched browser actions, and critical security patches for device pairing and webhook authentication. The release maintains macOS app version compatibility for future non-beta releases.

OpenClaw v2026.3.13 is a major release featuring significant UI/UX improvements across Android and iOS platforms, enhanced browser automation capabilities with Chrome DevTools MCP support, and numerous stability fixes. Key improvements include redesigned chat settings, first-run onboarding flows, batched browser actions, timezone configuration for Docker, and critical fixes for memory leaks, session management, and authentication handling across multiple integrations.

Claude Code v2.1.76 introduces MCP elicitation support for interactive structured input, new hooks for intercepting responses, and the `/effort` slash command for model effort control. The release includes significant bug fixes for deferred tools losing schemas after compaction, plan mode re-approval issues, voice mode input handling, and Remote Control session stability. Performance improvements include optimized worktree startup, better background agent handling, and improved model fallback notifications.

OpenClaw v2026.3.8-beta.1 introduces backup and verification CLI commands, configurable Talk mode silence timeouts, improved web search capabilities with Brave LLM Context support, macOS remote gateway token handling, and numerous fixes across Android, Telegram, Matrix, browser CDP, and model compatibility. Key enhancements include ACP provenance metadata tracking, TUI agent workspace inference, and WCAG AA-compliant light theme detection for terminals.

OpenClaw v2026.3.8 introduces significant enhancements across CLI backup operations, macOS onboarding with remote gateway tokens, configurable Talk mode silence timeouts, TUI workspace inference, Brave web search LLM context mode, and ACP provenance tracking. The release includes numerous fixes for platform-specific issues including macOS launchd restart recovery, Android permission cleanup, Telegram/Matrix DM routing deduplication, browser CDP normalization, and model compatibility updates for GPT-5.4.

OpenClaw v2026.3.12 introduces major UI/dashboard improvements with modular views, fast-mode toggles for OpenAI and Anthropic models, provider-plugin architecture for Ollama/vLLM/SGLang, Kubernetes deployment support, and subagent session control. The release includes 30+ fixes addressing security vulnerabilities (device pairing, plugin auto-load, Unicode obfuscation), model compatibility issues (Kimi Coding, OpenRouter, Moonshot), chat deduplication across platforms (Slack, Telegram, iMessage, Mattermost), and infrastructure stability (session discovery, plugin caching, Windows updates).

Google ADK Python v1.27.0 introduces major enhancements across core agent functionality, LLM model support, and observability. Key additions include A2A request interceptors, pluggable auth integrations, PDF support for Anthropic models, streaming capabilities, and new optimization tools. The release also expands tool ecosystems with Skills support, BigQuery enhancements, and improved human-in-the-loop interactions in the web interface.

This cookbook demonstrates Claude 3.7 Sonnet's extended thinking feature, which provides transparent step-by-step reasoning with budget management. Extended thinking enables Claude to show its internal reasoning process through thinking content blocks before delivering final answers. The guide covers setup, basic examples, streaming with thinking, token counting, redacted thinking, and error handling with practical Python code examples.

Claude 3.7 Sonnet may not make parallel tool calls by default even when enabled. This cookbook demonstrates a workaround using a "batch tool" meta-pattern that wraps multiple tool invocations, encouraging the model to call multiple tools simultaneously in a single response. By introducing a batch_tool that accepts an array of tool invocations, developers can improve latency and efficiency by processing multiple tool calls in parallel rather than sequentially.

Speculative prompt caching is a technique that reduces time-to-first-token (TTFT) by warming up Claude's cache while users are still formulating their queries. Instead of waiting until a user submits their question to load context into the cache, the system begins cache warming immediately when the user starts typing. This cookbook demonstrates the pattern using SQLite source code as example context, showing how to implement it with the Anthropic Python SDK using async operations and prompt caching controls.