Skip to content
Agents tracked: 258 Downloads (7d): 219M up 6.1% GitHub stars: 5.5M VS Code installs: 148M Releases (7d): 293 Agent status: 2 with issues Updated Oct 7, 2026
AgentGid

AgentScope changelog: what's new each month

Every stable AgentScope release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 12 months covered; the current month updates daily.

September 2026

2 releases: 2.0.8 → 2.0.9

September brought major expansions to AgentScope's multi-agent capabilities and voice support. The month saw the introduction of real-time voice agents, agent-to-agent communication, and new workflow orchestration tools for complex task management.

Highlights

  • Realtime voice agents now support OpenAI Realtime, Gemini Live, xAI Grok, and DashScope backends with local transport.
  • Agent-to-agent communication enabled through new A2AAgent class and team coordination via TeamPipeline for concurrent task execution.
  • New pipeline and workflow modules including GoalPipeline for goal-driven tasks and experimental Standard Operating Procedures for ordered milestone verification.
  • Terminal UI added for full-screen agent interaction and real-time voice sessions.
  • LLM-based reranking and agent-driven context compression improve retrieval and context management.

New

  • A2AAgent class enabling direct agent-to-agent communication.
  • GoalPipeline module for goal-driven agent workflows.
  • TeamPipeline class allowing leader agents to delegate tasks to concurrent member agents.
  • Standard Operating Procedures (experimental) for splitting tasks into ordered, verified milestones.
  • Terminal UI module (agentscope.tui) for full-screen chatting and real-time voice interaction.
  • AskUser external tool for agents to pose multiple-choice questions to users.
  • ModelRouterMiddleware with Jev-based classifier for intelligent model routing.
  • DingTalk knowledge base search tools for agents operating on DingTalk channels.

Improved

  • Realtime voice agent now supports additional models: OpenAI Realtime, Gemini Live, and xAI Grok Voice.
  • LLM-based reranking added to RAG retrieval to improve result quality.
  • Agent-driven context compression tool allows agents to actively manage their context window.
  • MCP support expanded with runtime headers in Docker workspaces and improved connection cleanup.
  • Web UI chat experience enhanced with session auto-naming, query caching, and improved dark mode.
  • Shell command permission checks tightened to better detect destructive git, sed, and tee usages.
  • Volcengine Ark chat model integrated.
  • MiniMax chat model (M3 by default) added.

Fixed

  • MCP connection cleanup improved to prevent resource leaks.
  • Shell permission confirmation now correctly triggered for previously undetected destructive commands using git, sed, and tee.

August 2026

3 releases: 2.0.6 → 2.0.7.post1

August brought workspace improvements, better multi-agent support, and new debugging tools to AgentScope. The month focused on agent isolation, skill management, and platform integrations.

Highlights

  • Workspaces can now be shared across different agents with per-agent skill isolation and session-level selection.
  • Added support for Feishu and Discord channels, Apple Container backend, and MCP skill hubs in agent services.
  • New terminal console for trying and debugging agents directly.
  • Web UI chat pages refactored with shadcn components.
  • Agent sessions are now properly isolated when switching between agents.

New

  • Support for Feishu and Discord channels in AgentScope.
  • Apple Container as a workspace backend option.
  • MCP and skill hubs/markets integration in agent services.
  • Terminal console for debugging and testing agents.
  • oncheckpermission middleware hook.
  • New pipeline module with GoalPipeline class.
  • Native multimodal tool output formatting.
  • Automatic final text summary generation after agents reach max iterations.

Improved

  • Workspace architecture refactored to support sharing across agents.
  • Skills can now be isolated per agent and selected on a per-session basis.
  • OpenAI AsyncClient is now reused across calls instead of creating a new instance per call.
  • Web UI chat pages redesigned with shadcn components.
  • Audio format now derived from media type automatically.
  • Session inbox payloads no longer strand until the next user turn.

Fixed

  • Removed unsupported $schema from Gemini tool parameters.
  • GitHub MCP environment inputs are now preserved in the hub.
  • Meta tool functionality fixed.
  • MoonshotChatModel streaming usage corrected.
  • Gemini tool-execution tokens excluded from output token count.
  • Session isolation when switching between agents in the Web UI.

July 2026

3 releases: 2.0.4 → 2.0.5

July brought significant expansions to AgentScope's capabilities, adding long-term memory systems, text-to-speech support, multiple workspace backends, and enhanced data handling. The month focused on making agents more autonomous, communicative, and capable of working with diverse data sources.

Highlights

  • Agent interruption with context handling allows better control over running agents.
  • Multiple long-term memory systems integrated for persistent agent learning and context retention.
  • Support for five different workspace backends including Kubernetes and sandboxing environments.
  • Text-to-speech APIs from OpenAI and DashScope enable agent voice output.
  • Agent-to-agent communication via AgentInvite for peer-to-peer team collaboration.

New

  • Agent interruption with context handling.
  • ReMe and agentic long-term memory integrations.
  • OpenAI TTS and DashScope CosyVoice TTS APIs.
  • Milvus Lite vector database support.
  • K8sWorkspace with Pod and PVC lifecycle management.
  • OpenSandbox, Daytona, and Bubblewrap workspace backends.
  • MongoDB and Elasticsearch vector store backends.
  • Excel and Word document parsers for RAG.

Improved

  • Session status endpoint added for better agent monitoring.
  • Structured output support for more predictable agent responses.
  • Environment information injection for contextual awareness.
  • PowerShell support on Windows for broader system integration.
  • SQLAlchemy-based storage for flexible data persistence.
  • Resource sharing across users as a group or organization.
  • Tool result formatting for Ollama now includes tool names.
  • Anthropic formatter improved by removing empty text blocks.

Fixed

  • Gemini schema generation containing null values.
  • Ollama tool result messages missing tool name information.
  • Anthropic formatter handling of empty text blocks.

June 2026

3 releases: 2.0.1 → 2.0.3

June brought major enhancements to AgentScope with native RAG capabilities, advanced memory management, and improved agent team features. The month focused on production-ready features for enterprise use cases including cost control and knowledge retrieval.

Highlights

  • Native RAG module with distributed, multi-tenant, and multi-session support for enhanced knowledge retrieval.
  • Long-term memory middleware powered by mem0 enables agents to retain persistent memory across sessions.
  • BudgetControlMiddleware for enforcing token budget limits and managing API costs.
  • Agent Team support with custom subagent templates and custom agent class options.
  • Streaming audio and live captions for omni models from DashScope and OpenAI.

New

  • Native RAG module supporting distributed, multi-tenant, and multi-session configurations.
  • mem0-backed long-term memory middleware for persistent agent memory across sessions.
  • BudgetControlMiddleware to enforce token budget limits and control API spending.
  • Streaming audio with live captions for DashScope and OpenAI omni models.
  • Custom subagent templates in agent service for flexible Agent Team configuration.
  • Custom agent class support within the agent service.
  • Onion-style middleware at tool level for flexible pre/post-processing around tool execution.
  • YAML configuration support for mainstream models with expanded model options.

Improved

  • Web UI now supports setting fallback models.
  • Model configuration enhanced with clientkwargs support.
  • Service infrastructure now provides extra tools and middlewares.
  • Graceful handling of invalid MCP config entries during workspace operations.
  • OpenAI-compatible structured output fallback to auto toolchoice when thinking mode rejects forced toolchoice.
  • Ollama embedding client creation per call to avoid event loop binding issues.
  • Redis session storage now preserves explicit session IDs.
  • Improved Anthropic formatter to drop thinking blocks without signature.

Fixed

  • Web UI example missing files restored.
  • DingTalk group QR code updated in documentation.
  • Qwen max 3.7 model ID corrected for DashScope.
  • Gemini function calls without an ID now handled correctly.
  • DingTalk integration documentation updated.

May 2026

3 releases: 1.0.20 → 2.0.0

May brought bug fixes and stability improvements across streaming, MCP tools, and model formatting, culminating in a major 2.0.0 release that refactors core messaging and tooling infrastructure.

Action needed
  • Evaluate, RAG, TTS, and realtime modules temporarily deprecated pending refactor.
  • Core Msg class structure simplified in 2.0.0.

Highlights

  • OpenAI response API support added for compatible providers.
  • MCP tool execution now includes timeout control and improved schema handling.
  • Major 2.0.0 release refactors message classes and tool architecture for better structure.
  • Fixed critical issues with streaming, formatter memory overhead, and token counting.
  • Thought signatures now supported for Gemini models.

New

  • MCP execution timeout configuration.
  • Extended reasoning effort options for OpenAI-compatible providers.
  • OpenAI response API support.
  • Thought signature support for Gemini models.
  • Basic permission classes framework.
  • New ToolBase class for builtin tools with comprehensive tests.
  • Skill loader class in toolkit module.

Improved

  • MCP tool input schema now preserves $defs and $ref properties.
  • Formatter memory overhead reduced and improved error handling.
  • TTS bug fixes with warning and retry logic added.
  • Anthropic formatter avoids emitting null text in tool results.
  • Tracing properly handles CancelledError when closing spans.
  • Milvus-lite version fixed.
  • Tool module refactored with new base class and toolkit logic.
  • Message type rules formalized with comprehensive tests.

Fixed

  • OpenAI response format errors during streaming now trigger fallback.
  • None input tokens in Anthropic stream message delta events handled.
  • Formatter crash when receiving None tool output.
  • DeepSeek null content rejection fixed.
  • Anthropic chat formatter tool result formatting.
  • MCP title duplication in schema $defs.

April 2026

2 releases: 1.0.19 → 1.0.19.post1

April brought multimedia support enhancements, new memory backend options, and improved error handling across multiple components. Several fixes addressed async operations and edge cases in tool handling.

Highlights

  • Local multimedia files can now be referenced using file:// paths in formatters.
  • TableStore memory backend is now supported for persistent agent memory.
  • MCP execution now includes configurable timeouts to prevent hanging operations.
  • Improved OpenAI streaming error handling with automatic fallback on response format issues.
  • Fixed missing await in DeepResearchAgent memory operations that could cause race conditions.

New

  • TableStore memory support for scalable persistent storage.
  • Execution timeout configuration for MCP tasks.
  • Local file:// path support in multimedia formatters.

Improved

  • OpenAI streaming now catches and handles response format errors gracefully with fallback options.
  • Better handling of missing 'type' field in OpenAI token counter tool parameter properties.

Fixed

  • Added missing await on memory.add call in DeepResearchAgent summarizing step.
  • Corrected field name from ReflectFailure.decomposesubtask.rephrasedplan to failedsubtask in deep research example.
  • OpenAI response format errors during streaming now trigger fallback instead of failing.

March 2026

2 releases: 1.0.17 → 1.0.18

March brought refinements to model integrations, tool handling, and agent capabilities. The month focused on improving compatibility with newer model APIs and expanding execution options.

Highlights

  • Tools now support background execution for improved performance
  • Enhanced tuner with prompt tuning and model selection capabilities
  • Better handling of local files without extensions in image conversion
  • Agents can now filter thinking content before broadcasting
  • Improved parameter override support in DashScope integration

New

  • Background execution support for tools
  • Prompt tuning capability in the tuner
  • Model selection enhancement in the tuner
  • Thinking content filtering for agents

Improved

  • DashScope model updated to use AioMultiModalConversation
  • Enhanced tool image URL conversion for local files without extensions
  • Parameter override handling in DashScope through generatekwargs repositioning
  • Home directory expansion in text file tools using tilde paths
  • Local path identification logic in formatter

Fixed

  • OpenAI model support for newer vllm reasoning field parsing
  • Double dot artifacts when saving base64 data with file extensions
  • Monitor configuration in tuner no longer overwrites monitor objects
  • HuggingFace endpoint compatibility
  • GitHub Actions compatibility with Node 24

February 2026

2 releases: 1.0.15 → 1.0.16

February brought realtime voice and video capabilities to AgentScope, along with new vector database support and several important bug fixes across core components.

Highlights

  • Realtime voice agent implementation enables live voice interactions.
  • OceanBaseStore added as a new vector database option for RAG workflows.
  • Video recording support added to realtime agent examples.
  • Multimodality parameter introduced for DashScope chat model.
  • Memory compression and serialization bugs fixed in ReActAgent and InMemoryMemory.

New

  • Realtime voice agent for live voice-based interactions.
  • OceanBaseStore vector database integration for RAG pipelines.
  • Video recording capability in realtime agent examples.
  • Logger and Tinker support in the tuner.
  • Multimodality parameter for DashScope chat model.
  • OpenJudge evaluation integration documentation.
  • Tool format support for OpenAI realtime model.
  • Pull request title validation for GitHub integration.

Improved

  • Websockets library now uses lazy import to reduce dependencies.
  • StreamPrintingMessages improved with sleep mechanism to prevent blocking.
  • Realtime event handling with typo fixes.
  • MCPToolFunction timeout argument handling.

Fixed

  • Memory compression bug in ReActAgent.
  • Serialization error in InMemoryMemory class.
  • Tool format compatibility with OpenAI realtime model.
  • Message consumption blocking in StreamPrintingMessages.
  • MCPToolFunction timeout argument behavior.
  • Tutorial bug and documentation issues.

January 2026

4 releases: 1.0.11 → 1.0.14

January brought expanded model support, improved memory capabilities, and new vector database integrations to AgentScope. The month focused on enhancing agent communication, data handling, and long-term memory features.

Highlights

  • Added support for Azure OpenAI API and Qwen-Deep-Research model integration
  • ReActAgent now supports automatic memory compression and relational database storage
  • New vector database options including MongoDB and AlibabaCloud MySQL for RAG workflows
  • A2A protocol for inter-agent communication now available with improved stability
  • Tool renaming capability and streaming mode tool parsing controls added

New

  • A2A protocol support for inter-agent communication
  • Azure OpenAI API compatibility
  • Qwen-Deep-Research model integration
  • MongoDB vector store for RAG module
  • AlibabaCloud MySQL vector store implementation
  • Excel reader support
  • Redis-based session storage
  • Custom tool renaming via funcname parameter

Improved

  • ReActAgent now supports automatic memory compression
  • Memory module extended to support various relational databases
  • Streaming mode now allows disabling tool use input parsing
  • Ollama model support improved for streaming output and mem0 integration
  • Gemini model tool handling with flattened JSON schema
  • A2A server implementation stabilized
  • Message formatting for OpenAI, Anthropic, and token counting fixed

Fixed

  • Werewolf game now broadcasts results correctly when ending at night phase
  • OpenAI formatter now correctly extracts image extensions from URLs with query parameters
  • Anthropic response content block type parsing logic error
  • Redis memory bugs including bytes data handling as message keys
  • Ollama event loop close issue in mem0 by ensuring single event loop usage
  • Mem0 compatibility issues with Ollama model

December 2025

2 releases: 1.0.9 → 1.0.10

December brought improvements to agent skills, model configuration, and deployment capabilities, with fixes to core functionality and documentation. The month focused on expanding integration options and refining existing features.

Action needed
  • clientargs argument renamed to clientkwargs in chat model classes

Highlights

  • Added support for Anthropic agent skills in ReActAgent and Toolkit class
  • Unified client configuration arguments across chat model classes for consistency
  • Added deployment example for routing agents as chat endpoints
  • Fixed critical bug in OpenAI chat model when handling empty delta responses
  • Improved tracing system to handle list inputs in agent message retrieval

New

  • Support for Anthropic agent skills in ReActAgent and Toolkit class
  • Video block support in DashScope formatter
  • Deployment example for routing agents as chat endpoints
  • List input support in getAgentMessages for tracing
  • Original function name tracking in RegisteredToolFunction

Improved

  • Unified clientkwargs argument across chat model classes to replace clientargs
  • Plan change hook now supports awaitable functions
  • Meta planner example refactored with PlanNotebook class
  • OpenTelemetry dependency versions updated

Fixed

  • Meta tool group deactivation failure in toolkit
  • Bug in OpenAI chat model when choice.delta is None
  • CreateAt timestamp format in configuration
  • TTS tutorial documentation error
  • Incorrect port number in tutorial documentation

November 2025

2 releases: 1.0.7 → 1.0.8

November brought stability improvements and refactoring of core agent logic, along with new memory and document reading capabilities for the RAG module.

Highlights

  • Added long-term memory support through REME integration.
  • Implemented word document reading in the RAG module.
  • Refactored ReActAgent response generation and structured output handling for better reliability.
  • Made configuration thread-safe using ContextVar.
  • Fixed audio format handling for Qwen-Omni models.

New

  • Long-term memory capability via REME integration.
  • Word document reader in RAG module.

Improved

  • ReActAgent response generation refactored for better performance.
  • Configuration system refactored to be thread-safe.
  • Separated enablemetatool and plannotebook argument logic in agents.
  • Structured output logic refactored within ReActAgent.
  • Tool choice set to 'required' when using structured output.

Fixed

  • Stream printing messages pipeline bug.
  • Planner state issue.
  • Qwen-Omni audio format requirement now handled differently from OpenAI models.
  • CancelError now properly raised during tool execution interruption.
  • Type substitution in MsgHub using Sequence for covariance.
  • Ollama embedding API integration.

October 2025

2 releases: 1.0.5 → 1.0.6

October brought stability improvements and expanded integration capabilities to AgentScope. The month focused on fixing critical bugs and adding support for new tools and platforms.

Highlights

  • Added async generator support to retrieve printing messages from agent instances
  • Implemented audio playback from URL functionality
  • Added support for uv package manager
  • Fixed critical issues in browser agent formatting and API key handling
  • Improved werewolves game example

New

  • Async generator interface for obtaining printing messages from agents
  • Audio playback from URL capability
  • uv package manager support
  • Demo example showing AgentScope connection to Alibaba Cloud
  • Tool agent examples in documentation

Improved

  • Werewolves game example in AgentScope
  • Tutorial refinements
  • README documentation with tool agent examples

Fixed

  • Missing await in browser agent's format function
  • Missing API key bug in DashScopeMultiModalEmbedding
  • Stream printing messages incorrectly printing generate response tool use messages
  • Plan notebook serialization bug
  • Missing default value None for memorytype in mem0record

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. AgentScope: pricing, features and alternatives · All changelogs