Multi-Agent Systems

Claude Code v2.1.178 introduces enhanced permission rules with parameter matching syntax, improved nested `.claude/` directory handling for skills and workflows, and refined auto mode with pre-launch subagent evaluation. The release includes better error messaging for Remote Control, improved `/doctor` output, and numerous bug fixes addressing subagent transcript display, authentication issues, and vim mode undo behavior.

Propolis is an autonomous QA platform that deploys swarms of browser agents to simulate user behavior, identify bugs, and generate e2e tests for web applications. The agents collaboratively explore websites, flag friction points, and propose tests that integrate into CI/CD pipelines. Available at $1000/month with flexible pricing options, it addresses the gap between deterministic testing and real-world usage coverage by treating agents as a canary group for quality assurance.

OpenClaw v2026.6.8-beta.1 introduces significant improvements to channel delivery (Telegram rich text, WhatsApp ACP bindings), agent and gateway recovery mechanisms, expanded provider/model support (GLM-5.2, Claude Haiku 4.5), enhanced usage hooks with footer rendering, and UI/mobile stability fixes. The release addresses 50+ issues across messaging, runtime reliability, authentication, memory management, and user interface components, with contributions from 30+ developers.

Claude Code v2.1.174 release includes UI improvements for model selection, fixes for environment variable inheritance in background sessions, corrections to Bedrock GovCloud region handling, and enhanced usage attribution tracking in VSCode. The update addresses multiple edge cases affecting model display, authentication, and performance, particularly for enterprise users with specialized billing configurations.

crewAI v1.14.7 introduces pluggable backends for memory, knowledge, RAG, and flow management, along with a native Snowflake Cortex LLM provider and chat API for conversational flows. The release includes significant bug fixes for checkpoint restoration, telemetry, file input handling, and security vulnerabilities. Performance improvements focus on lazy-loading docling imports to speed up crewai imports, while refactoring decouples conversational logic and splits flow.py into modular components.

CrewAI 1.14.7rc1 introduces runtime state management improvements with a new `reset_runtime_state` feature to prevent accumulated bus state, decouples conversation logic from runtime with a `conversational_definition`, and fixes critical issues around runtime state scoping per run to prevent growth and isolate concurrent runs. The release includes bug fixes for telemetry setup and event suppression, along with documentation updates for OpenTelemetry and refactoring improvements to flow condition evaluation and conversation routing.

crewAI releases version 1.14.7rc2, a release candidate that addresses a critical bug in snapshot restoration. The main fix gates the restore functionality on a flag to prevent live snapshots from being replayed during resume operations. This update includes documentation updates to the changelog and version information for the previous rc1 release.

Claude Code v2.1.172 introduces hierarchical sub-agent spawning (up to 5 levels deep), improves AWS region configuration handling for Bedrock, and adds a marketplace search feature. The release includes numerous bug fixes addressing context management, session stability, model selection, and UI responsiveness issues. Performance optimizations reduce CPU usage and improve message handling in long conversations.

This cookbook demonstrates two async multi-agent orchestration patterns using Claude: a fixed N-agent team with peer messaging through a shared hub, and dynamically spawned async subagents. Built on the Anthropic Python SDK and asyncio, it provides bare messaging and lifecycle mechanics without domain-specific tasks, allowing developers to see exactly which tools fire and in what order before adding their own tools and logic. The patterns run on any API key in under thirty seconds and include a message hub, messaging tools (send_message and wait_for_message), and a base agent loop that handles tool dispatching and inbox management.

crewAI v1.14.7a4 introduces runtime migration for @listen/@router decorators to read from FlowDefinition and adds pluggable default backends for memory, knowledge, RAG, and flow components. This release enhances the framework's flexibility by allowing developers to swap backend implementations without code changes. The update includes documentation updates and contributions from three core team members.

crewAI v1.14.7a3 is a patch release addressing critical security vulnerabilities and bug fixes. Key improvements include fixing the exposure of `ask_for_human_input` on the experimental AgentExecutor, resolving CVEs in dependencies (aiohttp, docling, docling-core, pip), and refactoring the `@start` decorator to read from FlowDefinition. This release maintains stability while enhancing security posture and internal architecture.

crewAI v1.14.7a2 introduces conversational flow traces support with enhanced LLM event transparency, including real finish_reason, sampling parameters, and response IDs. The release refactors the flow DSL into focused decorator modules and implements a chat API for conversational flows. Key improvements include type-aware route decorators, overridable locking backends, and flattened LiteLLM usage tracking, alongside new NVIDIA Nemotron documentation and monorepo deployment guides.

crewAI 1.14.7a1 introduces native Snowflake Cortex LLM provider support and crew trained agents file functionality. The release includes integration guides for Databricks and Snowflake, fixes critical CLI and file input issues, and improves import performance through lazy-loading. Key refactoring splits flow.py into modular components while addressing tool result handling and multi-source listener reliability.

Claude Code v2.1.161 introduces enhanced observability through OTEL metric labels, improves parallel tool execution by allowing independent failures, and fixes critical issues with clipboard handling, authentication policies, and background session management. The release includes UI/UX improvements for agent progress display, MCP connector management, and accessibility features like motion reduction support. Multiple regressions and edge cases are resolved, particularly around git worktrees, file editing in isolated workflows, and telemetry initialization.

LangChain version 1.3.3 introduces improvements to subagent run projection onto typed channels and enhances the HumanInTheLoopMiddleware with interrupt_mode and conditional predicate support. The release includes a dependency bump for langgraph to 1.2.4 and loosens the langgraph version constraint for better compatibility. These changes improve agent orchestration capabilities and middleware flexibility for human-in-the-loop workflows.

OpenClaw v2026.5.31-beta.1 introduces major stability improvements for agent runtimes, channel delivery, and plugin orchestration. Key enhancements include better recovery from interrupted tool calls, steadier multi-channel support (Telegram, WhatsApp, Slack, Discord, Teams, etc.), bounded timers and retries to prevent hangs, and new Skill Workshop features for governed skill creation. The release also externalizes Tokenjuice and GitHub Copilot as official plugins, adds Workboard orchestration primitives, and significantly improves CI/E2E testing infrastructure with better log bounding and failure reporting.

OpenClaw v2026.5.31-beta.2 introduces significant stability improvements for agent runtimes, channel delivery, and plugin management. Key enhancements include better recovery from interrupted tool calls, steadier multi-channel support (Telegram, WhatsApp, Slack, Discord, Teams, etc.), bounded timers and retries to prevent hangs, and new Skill Workshop orchestration features. The release also externalizes Tokenjuice and GitHub Copilot as official plugins, adds iOS push relay improvements, and strengthens CI/E2E testing with comprehensive logging caps and artifact validation.

OpenClaw v2026.5.30-beta.1 introduces significant stability improvements for agent runtimes, CLI tools, and multi-channel delivery across Telegram, WhatsApp, Slack, Discord, and other platforms. Key enhancements include better recovery from interrupted tool calls, bounded timers and retries to prevent hangs, and new orchestration primitives through Workboard and SecretRef integrations. The release externalizes Tokenjuice and GitHub Copilot as official plugins, adds Skill Workshop proposal lifecycle management, and improves iOS push relay reliability. Extensive fixes address cron migrations, provider media downloads, session cleanup, and CI/E2E test coverage routing.

OpenClaw v2026.5.28 delivers major stability and safety improvements across agent runtime recovery, channel delivery, mobile/chat surfaces, and provider coverage. Key enhancements include steadier subagent workspace separation, safer session identity handling across multiple channels (Matrix, Slack, Discord, Teams, etc.), broader mobile UI refresh with state preservation, stricter input validation, and expanded provider support (Claude Opus 4.8, Fal Krea, NVIDIA models). The release also improves CLI/auth paths, reduces repeated work in plugin hot paths, and strengthens E2E validation.

OpenClaw v2026.5.28-beta.4 delivers significant stability and safety improvements across agent runtime recovery, channel delivery, mobile/chat surfaces, and provider coverage. Key enhancements include steadier subagent workspace separation, safer session identity handling across multiple platforms (Matrix, Slack, Discord, Teams, WhatsApp, Telegram, iMessage), broader mobile UI refresh, stricter input validation, expanded provider support (Claude Opus 4.8, Fal Krea, NVIDIA models), and faster CLI/auth failure detection with clearer recovery paths.