openJiuwen changelog: what's new each month
Every stable openJiuwen release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 5 months covered; the current month updates daily.
October 2026
1 release: 0.1.20
October brought significant collaboration and multimodal capabilities to openJiuwen, enabling teams to work together on complex tasks and interact with agents through voice and audio in real time.
Highlights
- Team Organization for multi-expert-team collaboration on cross-team tasks.
- Scheduled dispatch mode with verifier and challenger review roles for Agent Teams.
- Full-duplex voice and microphone support for real-time audio interaction.
- Computer Use operations now triggerable via voice commands.
- Third-party expert integration including Claude Code, Codex, and DSH.
New
- Team Organization feature for coordinating multiple long-lived professional teams.
- Scheduled dispatch mode with binary voting and traceable review for teams.
- Timeout escalation for team decision-making.
- Multimodal full-duplex interaction with microphone and voice.
- Real-time audio-stream processing capabilities.
- Built-in openJiuwen Skill support for third-party agents.
- Third-party expert harnesses in single-agent and cluster modes.
Improved
- Swarmflow execution engine enhanced for Design mode with expert routing.
- Canvas-based multimodal creation pipeline with improved expert routing.
- Third-party expert streaming through chat.send interface.
September 2026
2 releases: 0.1.18 → 0.1.19
September brought major advances in agent reliability, orchestration, and autonomous evolution. The month focused on making multi-agent systems more robust, observable, and capable of self-improvement.
Highlights
- Agent Teams ReliabilityRail detects and handles common failures like model errors, tool errors, timeouts, and rate limiting with configurable escalation.
- SwarmFlow adds budget controls and workflow nesting for composable, constrained orchestration of agent teams.
- TTSE dual-track self-evolution framework enables agents to autonomously improve by learning from environment and task experience.
- Web-based observability dashboard provides visual monitoring of agent execution and swarm behavior.
- Structured execution traces capture reasoning, tool decisions, and outcomes for session-level analysis.
New
- ReliabilityRail detectors for repeat-tool, model-error, tool-error, output-length, compaction, and ping-pong failures with leader escalation.
- Retry and user-guidance handling for transient failures, rate limiting, and service unavailability.
- Budget operators and nested workflow support for orchestration constraints.
- Batch Skill graph operations with atomic version publishing.
- AgentTemplate and Plugin system with natural-language, form-based, and package-upload creation.
- Proactive Context Service for multi-source context construction and injection.
- Trajectory CRUD APIs with LoRA weight version management and hot-loading.
- Skill package installation and execution in Agent Core.
Improved
- Online reinforcement learning workflows with explicit training start/stop controls and fallback mechanisms.
- Agent-Team prompts, identities, and tool descriptions can now evolve with execution experience.
- Trajectory-driven capability composition discovery identifies high-quality skill combinations.
- Permission Engine architecture improved separation of policy and enforcement.
- Runtime control and budget governance with TUI commands for operational workflows.
- Swarm evolution framework expanded for autonomous capability improvement.
- Structured JSON collection of execution traces for comprehensive debugging.
August 2026
1 release: 0.1.17
August brought significant expansions to openJiuwen's multi-agent capabilities and session management. The release focused on integrating external agents, adding persistent goal tracking, and extending execution control across teams and workflows.
Highlights
- Teams can now integrate third-party agents alongside native members for more flexible agent composition.
- DeepAgent sessions now maintain goals across reconnects with lifecycle management commands.
- Multi-agent browser execution lets team members run isolated WebAgent sessions concurrently.
- Harnesses support multimodal inputs including images and video through Skill-Omni integration.
- Prefix caching hooks reduce token costs and latency for repeated model requests.
New
- Third-party agent integration for AgentTeam.
- Session-scoped goals with set, pause, resume, clear, and assess-progress lifecycle commands.
- Human-in-the-loop nodes in SwarmFlow.
- Stateful agents and session workers for SwarmFlow.
- Budget constraints in SwarmFlow runtime.
- Multimodal input support (image and video) for Harnesses via Skill-Omni.
- Explicit and dynamic prefix caching for model requests.
- OAuth account authentication for third-party model-service plans.
Improved
- SwarmFlow operators now support branch and operator-level suspend, resume, and cancel controls.
- Multi-agent WebAgent execution with headless support and resource cleanup.
- WebAgent team members can run in isolated concurrent browser sessions.
Fixed
- Deleting a team before it has been built no longer causes errors.
July 2026
1 release: 0.1.16
July brought significant improvements to team-based agent operations, with new observability capabilities, permission controls, and memory management features. The release focused on making multi-agent workflows more reliable and traceable.
Highlights
- Full observability for Agent Teams through OpenTelemetry trace export for end-to-end monitoring.
- Team-level permission management system for controlling access within Agent Teams.
- Extended memory storage with Lakeon backend support, including versioning and branch management.
- Unified Python reporting for both high-code and low-code OTEL data to external systems.
- Worktree isolation for teammates now aligned with AgentTool workflow standards.
New
- OpenTelemetry trace export for Agent Teams observability
- Team permission control system
- Lakeon memory storage backend with versioning and branch support
- Unified OTEL-based data reporting for external operations and maintenance systems
Improved
- Worktree isolation for teammates aligned with AgentTool flow
- MemoryProvider extended to support Lakeon backend
Fixed
- Agent Team members now properly learn from the Context Evolver algorithm
- Human members no longer complete tasks without user input in HITT mode
- Memory tools remain available after multiple dialogue rounds in agent team mode
June 2026
15 releases: 0.1.1 → 0.1.15
June brought major expansions to openJiuwen's agent capabilities, including multi-agent collaboration frameworks, advanced memory systems, and self-evolution features. The month saw numerous refinements across agent types, retrieval systems, and infrastructure components.
Highlights
- Introduced Agent Teams framework with leader-teammate collaboration, enabling structured multi-agent task decomposition and assignment.
- Launched self-evolution framework allowing agents, tools, and context to automatically optimize themselves based on experience.
- Added semantic and episodic memory support with graph-based implementation and query rewrite capabilities for improved retrieval.
- Implemented NL2Agent for low-code agent generation from natural language descriptions.
- Introduced human participation in agent teams as first-class citizens with dedicated human agent roles.
New
- ReActAgent and WorkflowAgent with multi-turn reasoning, self-correction, and complex task automation.
- Graph Knowledge Base implementation based on GeAR paper for enhanced reasoning and deep context retrieval.
- Dual-engine retrieval supporting Vector, Sparse, Hybrid, Graph-based, and Agentic methods.
- Summary memory capabilities with built-in compression and offloading to handle information overload.
- Guardrail framework for embedding security inspections within agent lifecycle.
- Pluggable session checkpointing with local and Redis-based storage for agent recovery.
- DeepAgents browser automation with profile persistence for end-to-end task automation.
- IntelliRouter LLM proxy for routing across multiple backend MaaS endpoints.
Improved
- Context processing with configurable strategies and extensible compression methods.
- Workflow component selection now correctly handles nested data types with comparison operators.
- Memory retrieval prompts are now configurable instead of hard-coded.
- Context expansion with three memory algorithms (ACE, RB, ReMe) for better knowledge retrieval.
- Enhanced Harness development framework with middleware extension and multi-task scheduling.
- Real-time intelligent compression of long tool messages while retaining key information.
- Team skill support for enhanced skill sharing and scheduling in multi-agent collaboration.
- Tool information method refactored for better architecture separation.
Fixed
- ReActAgent tool invocation failures and memory scope configuration not being applied.
- WorkflowAgent intent recognition logic failures.
- Milvus vector store collection loading safety check before attempting access.
- Branch component selection with nested data types and comparison operators.
- Sub-workflow execution when containing loop components producing no output.
- End component streaming with template causing runtime failures and output duplication.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. openJiuwen: pricing, features and alternatives · All changelogs