AgentScope Community Office Hours

[DavdGao]

We are launching the AgentScope Community Biweekly Meeting, welcoming all community members to participate in discussions, ask questions, and share insights.

Meeting Schedule: Every two weeks - specific dates and times will be announced in advance in this issue
Meeting Link: Will be updated before each session
During the biweekly meetings, we will:

• Share recent progress and updates in the AgentScope ecosystem
• Discuss upcoming development plans and roadmap
• Address community questions and gather feedback

We will continuously update this issue with:

• 📅 Biweekly meeting announcements and schedules
• 🔗 Meeting links for each session
• 📝 Meeting agendas and discussion topics

Stay tuned to this issue and join us in the AgentScope community discussions!

[DavdGao]

📅 First Biweekly Meeting

Date & Time: January 19, 2026, 19:30-20:30 (Beijing Time / UTC+8)

Meeting Link: https://meeting.dingtalk.com/j/K7mU71JQa0A

📝 Agenda

AgentScope Core

• Memory
• Agent skill
• Voice agent
• Training

AgentScope-Runtime

• Deployment strategies
• Protocol standards
• Tool sandbox implementation

Alias

• Preview version release
• Version compatibility
• QA-agent improvements
• Input data profiling

AgentScope-Java

• Agent skill
• Live API
• HITL

AgentScope-Studio

• UI/UX optimization
• Voice integration
• Evaluation features

---

All community members are welcome to join and participate in the discussion!

[ingto]

有录制嘛？想来，但跟其他会议时间冲突了

[chickenlj]

Yes. It will be recorded.

[wang123123wjl]

怎么一打开就显示结束了

[gitriver]

录制内容在哪里看

[DavdGao]

Meeting Minutes - AgentScope Biweekly Sync

Date: January 19, 2026

AgentScope Ecosystem Overview

The AgentScope ecosystem consists of:

• AgentScope (Python/Java): Core agent development framework
• Runtime: Secure sandbox and deployment support
• Studio: Visual development and debugging tool
• Spark Design: Frontend UI component library
• AgentScope-samples: Ready-to-use agent applications
• Integrated frameworks: Reme (memory management), OpenJudge (evaluation), Trinity (agent RL training)

Module Updates

1. AgentScope-Python (v1.0.12)

Recent Updates:

• Added database support: in-memory, relational DB, Redis, and NoSQL
• Implemented memory compression with new mark interface
• Agent skill registration via register_agent_skill interface

Planned:

• Expand NoSQL support
• Memory offloading mechanism for short-term memory
• Skill directory monitoring and dynamic management
• Human-in-the-loop functionality for tool execution control

2. Tuner (Fine-tuning Module)

Key Features:

• Zero code modification required for fine-tuning
• Low learning curve - only requires workflow, reward function, and dataset
• Local debugging without GPU; remote GPU support via Tinker or Alibaba Cloud DLC
• Multi-agent fine-tuning support (e.g., 30B model for Werewolf game)

Upcoming:

• Continuous fine-tuning based on application history
• Enhanced reward functions with auto-generation
• Improved data augmentation for few-shot training
• Special Event: Tuner livestream on Wednesday, 7:00 PM

3. Voice Agent

Current Support:

• ReactAgent voice output via speech field
• TTS integration supporting DashScope, OpenAI, and Germany TTS
• Realtime/non-realtime and streaming/non-streaming modes
• WebSocket long connection with low-latency callback

Roadmap:

• Real-time voice agent with interruption support
• Integration with DashScope Realtime API, GPT Realtime, and Gemini Live API

4. Runtime (v1.0.4/v1.0.5)

Updates:

• Cloud-native deployment: Alibaba Cloud FC, PAI, K8S KNative
• Framework-agnostic: supports Microsoft Agent Framework, LangGraph, CrewAI (in progress)
• Google A2A registry and AGUI protocol support
• Async SDK and non-Docker options (gVisor, BoxLite)

Future Plans:

• Sandbox persistence and lifecycle management
• FastAPI ecosystem integration
• Enhanced multi-tenant support
• Third-party sandbox integration (E2B in development)

5. Studio (v1.0.7)

Major Features:

• Dual-view message display: timeline view and aggregated reply view
• React agent intermediate state visualization
• OpenTelemetry integration for full-chain tracing
• Docker deployment support

Roadmap:

• Real-time voice playback
• Evaluation module visualization with benchmark analysis
• Solution comparison functionality
• Integration with runtime Web UI for remote conversation management

6. AgentScope-Java (v1.0.7)

Core Updates:

• Agent skill repository abstraction for dynamic loading
• Human-in-the-loop support with pause/resume control
• Streaming tool call output (v1.0.6+)

Comparison with Spring Alibaba:

• AgentScope: React agent pattern with autonomous decision-making
• Spring Alibaba: Process orchestration (graph/workflow) focused

7. Alias (Universal Agent)

Project Goals:

• Ready-to-use agent application
• Development template for community

Five Modes:

• General task decomposition
• Browser interaction
• Research & financial analysis
• Data science

Recent Work:

• Compatibility upgrade for AgentScope v1.0.7-1.1.0+
• Agent runtime integration
• Data type classification and profiling development

Plans:

• Demo house with replayable examples
• Open input with daily usage quotas
• Multimodal support migration to community homepage

Join us for the next meeting on February 2, 2026!
Feel free to share feedback and suggestions in this discussion thread.

[DavdGao]

📅 Second Biweekly Meeting

Date & Time: February 2, 2026, 19:30-20:30 (Beijing Time / UTC+8)

Meeting Link: https://meeting.dingtalk.com/j/Ll96r3yPZE6

📝 Agenda

• TuFt(A new repo for lora-based tuning service)

• Anybox

• AgentScope

  • Human-in-the-Loop
  • Realtime Agent
  • A2UI

• AgentScope-Runtime

  • v1.1.0 Update

• Alias

  • Support Agent Configuration

[DavdGao]

Meeting Minutes - AgentScope Technical Progress

Date: February 2, 2026

---

TuFT

Multi-tenant fine-tuning for local LLMs with Tinker-compatible API

• Based on Trinity-RFT
• Unified backend GPU resource scheduling with support for Tinker framework compatibility
• Integrated LoRA adapter as a unified conversion library

Anybox (Planned)

A scalable system for building and running agent environments. It supports declarative environment definitions and provides full lifecycle management from build to scale.

AnyBox will be open-sourced soon.

AgentScope

RealtimeAgent

The PR of realtime agent:

• Added RealtimeAgent module with DashScope and OpenAI API support
• Implemented WebSocket-based real-time text/audio input and tool invocation
• Multi-model theory validation

Human-in-the-Loop

• New generate_reason interface in the Msg class
• Registration interface in Toolkit enabling external tool integration
• Allow to skip the reasoning pharse in the ReActAgent class

A2UI (Experimental Phase)

Implementation in our example

• Adopted skill-driven dynamic system prompt generation
• Provided offline model inference with speed improvements
• Calculated and introduced HUI v0.9+ similarity metrics

AgentScope-Runtime

1.1.0 Preview

Key Features:

• New sandbox system supporting local/remote data source connectivity
• Integrated automatic recovery mechanism with self-healing capabilities
• Native FastAPI support with lifespan and HTTP endpoint compatibility
• Enhanced multi-user resource isolation and stability

Value: Build reliable, autonomous runtime environment for Agents

Connection Rule Engine (Fully Integrated)

Deliverables:

• Developed DISCORD, $T$T, QQ and other cross-platform connection tools
• Task-role and timeline-based trigger support
• Optimized enterprise-level Agent workflow
• Automated arrival scene closure

Value: From experimental validation to real business scenarios

Alias

Data Processing Skills (In Progress - Medium Priority)

Before:

• Manual CSV data parsing with simple queries only

After:

• Automatic data type inference with format standardization templates
• Deduplication, fast search, and missing value handling
• Enhanced intelligent data preprocessing capabilities

Impact: AI data processing reduced by thousands of efforts

Configurable Agents (First-Class Priority)

Progress:

• Passed QA Agent and Deep Research Agent testing
• Exposed system prompt, model, and tool list parameters
• Multi-question expansion support (.aiplan Notebook)
• Tool workflow standardized and sealed

Impact: Development threshold reduced by half, accelerating productization