Skip to content
Agents tracked: 501 Releases (7d): 777 Agent status: no reported issuesSince last limit reset: Claude 10d · Codex 8d For AI agents: Markdown Updated Nov 21, 2026

BeeAI Framework changelog: what's new each month

Every stable BeeAI Framework 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.

November 2026 · so far

2 releases: 0.1.86 → 0.1.32

November brought dependency management improvements and several tool updates to the BeeAI Framework. The month focused on better integration configuration and removal of deprecated components.

Action needed
  • GoogleSearchTool has been removed; migrate to alternative search solutions.
  • AgentStack integration has been deprecated.
  • Transformers extra no longer automatically installs torch, peft, or outlines; install separately if needed.

Highlights

  • GoogleSearchTool has been removed due to API changes.
  • MiniMax adapter now supports regional endpoint selection for China.
  • Tool execution now supports per-attempt timeouts.
  • Memory improvements now include tool calls and results in summarization.
  • MCP image and audio payloads are kept out of model text for better efficiency.

New

  • Per-attempt timeout configuration for tool runs.
  • MiniMax regional (CN) endpoint selection in adapters.
  • Cached token usage reporting in the LangChain chat adapter.

Improved

  • MiniMax text models now default to M3.1 Flash.
  • Memory summarization now includes tool calls and results in prompts.
  • SummarizeMemory prompts enriched with tool execution context.
  • OpenAPITool parameters are now routed by their location.
  • JSON Schema property handling accepts values without explicit type definitions.
  • Vercel chat model now correctly passes maxTokens as maxOutputTokens.

Fixed

  • MCP image and audio payloads no longer included in model text.
  • Token usage caching now properly reported in LangChain adapter.
  • Tool parameter routing improved for OpenAPI tools.

September 2026

3 releases: 0.1.84 → 0.1.85

September brought improvements to memory persistence, adapter reliability, and framework infrastructure. The month focused on fixing serialization issues and enhancing compatibility across multiple AI model providers.

Highlights

  • Added DakeraMemory for persistent cross-session agent memory, enabling agents to retain information across separate sessions.
  • Ported the Runnable interface to TypeScript, extending core framework capabilities to TypeScript users.
  • Fixed critical serialization issues affecting six chat adapters and nested dictionary handling.
  • Improved stability when streaming calls fail in Vercel chat adapter to prevent process crashes.
  • Enhanced Amazon Bedrock adapter compatibility with ChatModel.fromName and clone methods.

New

  • DakeraMemory for persistent cross-session agent memory
  • Runnable interface now available in TypeScript
  • OpenAI server support
  • Typed schema generation for keyword-only parameters in tools

Improved

  • Amazon Bedrock adapter now supports ChatModel.fromName and clone operations
  • Stopped sending unnecessary default temperature values to OpenAI GPT-5.x and GPT-6 on Amazon Bedrock
  • LangChain image content is now preserved in adapters
  • File and shell tools now preserve cache and middleware when cloned
  • FileCache.set now properly awaits persistence operations
  • Agent task template rendering now includes backstory for model context
  • AbortController clone now properly carries abort state
  • Settings dict is copied instead of aliased to prevent unintended mutations

Fixed

  • Fixed cosine similarity matrix column count comparison
  • Fixed serialization of nested dictionary values
  • Added missing serializer registration to six chat adapters
  • Vercel chat adapter no longer crashes when streamed calls fail
  • Fixed Python stream errors and cancellation propagation
  • Strengthened API key comparison using constant-time operations

August 2026

1 release: 0.1.83

August brought improvements to error handling and streaming capabilities across the framework, with better support for the latest MCP protocol features.

Highlights

  • Error context now propagates through MCP tool errors for better debugging.
  • Streaming support added to WatsonxOrchestrateAgent.
  • Framework now supports the Responses API text content parts in adapters.
  • Improved error propagation from emitters to prevent silent failures.

New

  • Streaming support for WatsonxOrchestrateAgent
  • Support for Responses API text content parts in adapters

Improved

  • Error context propagation through meta on MCP tool errors
  • Error propagation from emitters
  • Default value added for maxrequestbodysize field following MCP 1.29.0
  • Group ID handling when cloning emitters

July 2026

2 releases: 0.1.30 → 0.1.82

July brought security improvements, dependency upgrades, and refinements to core components across the BeeAI Framework. Multiple fixes address serialization safety, adapter behavior, and token management.

Action needed
  • Function deserialization is disabled by default; enable explicitly if needed.
  • MiniMax default model changed from previous version to M3.

Highlights

  • Function deserialization is now disabled by default for improved security.
  • LangChain adapter packages upgraded to version 1.x.
  • MiniMax default model upgraded to M3 with Python and TypeScript support.
  • Token capacity is now enforced in TokenMemory to prevent overflow.
  • Event listener matching improved for compiled regex patterns.

New

  • Clear error messages when attempting to use Groq embeddings, which are unsupported by the Groq API.
  • Opt-in function deserialization in the serializer component.

Improved

  • LangChain embedding model adapter no longer uses hardcoded model IDs.
  • Event emitter off() method now correctly removes only matching listeners when using compiled regex.
  • Event emitter regex matching now works anywhere in the event path, not just at the start.
  • OpenAIServer now defaults to loopback binding with warnings for unauthenticated connections.
  • Data URIs in adapters are now inlined as base64.
  • TokenMemory now enforces token capacity limits.
  • Wikipedia dependency updated to version 2.5.0.

Fixed

  • Cache entries are now cleared when a ChatModelToolCallError occurs.
  • Vulnerable transitive TypeScript dependencies have been pinned.
  • Python test errors related to pyrefly and ruff have been resolved.

May 2026

3 releases: 0.1.80 → 0.1.81

May brought improvements to agent configuration stability, MCP tool handling, and platform compatibility. The framework now supports Python 3.14 and adds new adapters for exposing agents.

Highlights

  • Added Zed ACP adapter for exposing agents over stdio
  • Made PromptTemplate configuration use deep copy to prevent state sharing between agents
  • Added Python 3.14 support
  • Made MCP tool excludenone parameter configurable
  • Improved tool message handling and LiteLLM text preservation with tool calls

New

  • Zed ACP adapter for serving agents over stdio
  • MCP output types support
  • Python 3.14 compatibility
  • Configurable excludenone in MCPTool parameters
  • Backend switching capability for DuckDuckGo Search Tool

Improved

  • PromptTemplate config now uses deep copy to prevent mutable state sharing across agents
  • Tool message handling now ignores empty messages
  • LiteLLM text is preserved when tool calls are present
  • Class name validation added in dynamic adapters
  • Dependencies resolution improved

Fixed

  • Mutable state sharing in PromptTemplate configuration across agents
  • Empty tool messages causing issues in processing
  • LiteLLM text loss during tool call operations

March 2026

2 releases: 0.1.28 → 0.1.79

March brought security updates to the BeeAI Framework, addressing vulnerabilities through dependency upgrades.

Highlights

  • Security vulnerabilities fixed through updated dependencies

Fixed

  • Security issues addressed via dependency updates

February 2026

3 releases: 0.1.77 → 0.1.78

February brought dependency updates and SDK improvements to BeeAI Framework, including a migration to VercelAI SDK v6 and security fixes for known vulnerabilities.

Highlights

  • Migrated backend to VercelAI SDK v6 for improved compatibility and features.
  • Updated LangChain and dependencies to address CVE security issues.
  • Extended VectorStore and VectorStoreSearchTool with additional parameters.
  • Improved error handling for Groq integration.

New

  • Extended VectorStore and VectorStoreSearchTool parameters for greater flexibility.
  • Backend now sets finish reason for manual tool calls.

Improved

  • Migrated to VercelAI SDK v6 for better SDK alignment.
  • Enhanced error handling for Groq provider.
  • Updated agentstack-sdk to latest version.

Fixed

  • Resolved CVE security vulnerabilities in LangChain and related dependencies.

January 2026

5 releases: 0.1.75 → 0.1.26

January brought improved reliability and error handling across the BeeAI Framework, with better recovery from malformed tool calls and new retrieval-augmented generation capabilities through Agent Stack integration.

Highlights

  • Auto-recovery from malformed tool calls and empty responses
  • Agent Stack RAG capabilities with vector store and embedding model support
  • Middleware support added to ChatModel class for extended functionality
  • Tool call ID inference for Watsonx chat model
  • Enhanced event propagation in agents

New

  • Agent Stack RAG capabilities including AgentStackVectorStore and AgentStackEmbeddingModel
  • Middleware support in the ChatModel class
  • isEmpty helper for StreamToolCall

Improved

  • Auto-recovery from malformed tool calls and empty responses
  • Error recovery handling for malformed tool calls
  • Tool call ID inference for Watsonx chat model
  • Event propagation in agents
  • Handling of empty chunks and invalid responses from Watsonx
  • AgentStackChatModel settings propagation in adapters

Fixed

  • Malformed tool call handling
  • Empty response handling
  • Missing tool call IDs for Watsonx chat model
  • Empty chunks and invalid responses from Watsonx

December 2025

6 releases: 0.1.71 → 0.1.23

December brought significant enhancements to agent capabilities and adapter integrations, with new agents introduced and several improvements to prompt caching, streaming, and middleware handling. The framework expanded its tooling options and added better observability through cost and token tracking.

Action needed
  • ToolCallingAgent deprecated; migrate to alternative agents.
  • Support for top-level union operators in Bedrock adapter removed.

Highlights

  • New RequirementAgent for handling complex requirements with prompt caching support.
  • New LiteAgent and ThinkTool added for expanded agent functionality.
  • Cost and token usage information now included in agent output.
  • Tool call streaming support added to the backend.
  • Handoff Tool introduced for agent-to-agent communication.

New

  • RequirementAgent with prompt caching strategy for Bedrock.
  • LiteAgent for lightweight agent operations.
  • ThinkTool for agent reasoning capabilities.
  • Handoff Tool for agent transitions.
  • GlobalTrajectoryMiddleware for execution tracking.
  • Tool call streaming support in ChatModel.
  • Priority-based callback registration in Emitter.
  • Update method for PromptTemplate class.

Improved

  • RequirementAgent now handles prompt caching edge cases with Bedrock.
  • AgentStack SDK updated with default parameters for factory method.
  • RunContext events extended to support execution interception.
  • Change propagation improved in AgentStackChatModel.
  • AgentStack form extension updated.
  • GlobalTrajectoryMiddleware now prevents event duplication.
  • Performance improvements in agent execution.
  • Middleware property added to BaseAgent for better customization.

Fixed

  • Incorrect import in GlobalTrajectoryMiddleware resolved.
  • Bedrock adapter now handles top-level union operators correctly.
  • AgentStack server connection issues fixed.

November 2025

9 releases: 0.1.62 → 0.1.21

November brought numerous stability improvements and expanded adapter capabilities to BeeAI Framework. The month focused on fixing streaming issues, enhancing tool handling, and adding support for new AI providers and deployment patterns.

Action needed
  • BeeAIPlatform adapter renamed to AgentStack

Highlights

  • Auto-fix for malformed tool calls and improved streaming with incomplete Unicode handling
  • New support for Qwen and DeepSeek providers alongside existing models
  • Human-in-the-loop support through ioconfirm adapter and health check endpoints for A2A
  • Nested trajectory support and lazy loading for AgentStack adapters
  • OpenAPITool can now be created directly from a URL

New

  • Create OpenAPITool directly from URL instead of manual specification
  • Qwen and DeepSeek AI provider support
  • Health check support for A2A adapter
  • ioconfirm adapter for human-in-the-loop workflows
  • Custom message formatter for GlobalTrajectoryMiddleware
  • Nested trajectory support for AgentStack
  • Host parameter for A2A configuration

Improved

  • Auto-fix for malformed tool calls during execution
  • Tool cloning mechanism for better reliability
  • Error handling for incomplete Unicode chunks in streaming responses
  • Error handling improvements for Groq provider
  • HandoffTool sanitization and error handling
  • Memory propagation to context in ReactAgent
  • Type annotations for ConditionalRequirement
  • StreamToolCallMiddleware now propagates updates without streaming

Fixed

  • OpenAI Chat model no longer imports unnecessary dependencies
  • HandoffTool cloning now works correctly
  • ChatModel and EmbeddingModel cloning issues resolved
  • Streamed tool calls without IDs now handled properly
  • Content extraction from corrupted chat output chunks
  • GlobalTrajectoryMiddleware now emits events correctly

October 2025

11 releases: 0.1.52 → 0.1.61

October brought refinements across adapters, agents, and backend systems, with focus on streaming improvements, platform integrations, and better error handling. The month culminated in a platform rebrand and expanded configuration options.

Action needed
  • BeeAI Platform adapter renamed to AgentStack; update references in configuration.

Highlights

  • RequirementAgent now streams final answers and properly propagates retry parameters to the LLM.
  • Added tool choice mapping across all platform providers for consistent behavior.
  • Improved streaming by preventing token duplication and adding smart parsing for MCP tool output.
  • BeeAI Platform adapter renamed to AgentStack with new settings extensions.
  • Backend now automatically retries on empty or malformed responses.

New

  • Dynamic registration of BeeAI Platform extensions through adapters.
  • Tool metadata now included in BeeAIPlatformServer.
  • Listener priority settings for the emitter system.
  • OpenAI Chat Completion and Responses serve module added.
  • A2A server context support for better request handling.
  • Settings extensions for AgentStack agents.
  • Smart parsing for MCP tool output.
  • Support for additional message content types in the backend.

Improved

  • RequirementAgent now streams final answers instead of waiting for completion.
  • Tool output properly serialized when serving via MCP.
  • Emitter cleanups now handled correctly when cloning.
  • HandoffTool accepts any runnable for greater flexibility.
  • Default timeout added for A2A client connections.
  • Complex schema processing improved in internals.
  • Provider-specific settings and parameters better propagated in platform chat models.
  • Error context serialization improved for better debugging.

Fixed

  • Tool visibility regression in RequirementAgent.
  • Double-escaped tool calls now properly handled.
  • Commas in location names in the OpenMeteo tool.
  • Unsupported JSON Schema keywords (Groq) handled gracefully.
  • Missing messages in BeeAI Platform adapter.
  • Streamed tool calls with streamable=True now correctly parsed.

September 2025

12 releases: 0.1.41 → 0.1.51

September brought major API refactoring to align agents and chat models with a unified runnable pattern, along with expanded platform integrations and improved tool handling across multiple backends.

Action needed
  • Agents now require a positional input argument (str or list[AnyMessage]) as first parameter.
  • Agent options moved to AgentOptions typed dictionary as keyword arguments.
  • Return types renamed: TooCallingAgentRunOutput → TooCallingAgentOutput, ReActAgentRunOutput → ReActAgentOutput, RequirementsAgentRunOutput → RequirementsAgentOutput.
  • ChatModel.create() renamed to ChatModel.run().

Highlights

  • Agents and ChatModel now follow a consistent runnable interface with simplified input/output contracts.
  • Added support for parallel agent runs and multiple protocol options in the BeeAI Platform server.
  • Extended adapter support for LangChain, Transformers, Google VertexAI, and IBM Watsonx backends.
  • RequirementAgent graduated from experimental status with improved documentation and case-insensitive permission handling.
  • Fixed MCP tool hanging issues and improved error propagation across tool integrations.

New

  • Parallel agent execution support in BeeAI Platform.
  • A2AAgent and A2AServer for agent-to-agent communication in TypeScript.
  • LangChain ChatModel integration adapter.
  • Transformers ChatModel adapter for local model support.
  • Multi-protocol support for running the A2A server.
  • Streamable HTTP support in MCP.
  • Agent support in MCPServer.
  • Extensions support in BeeAIPlatformAgent client.

Improved

  • PromptTemplate simplified and extended.
  • Event registration and deregistration in emitter improved.
  • MCP server outputs flattened for cleaner results.
  • RequirementAgent inline documentation enhanced.
  • Yes/no responses in AskPermissionRequirement now case-insensitive.
  • Tool choice support now included when cloning backends.
  • Response format serialization for structured outputs.
  • Nested initialization support for BeeAIPlatformChatModel.

Fixed

  • RequirementAgent now handles empty message lists correctly.
  • MCP tool no longer hangs when server is not running.
  • MCP errors now propagate correctly through tools.
  • TypeError during retry logging fixed.
  • Form cancellation in BeeAIPlatformServer now handled properly.
  • Google VertexAI parameters correctly propagated in adapters.

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