MateClaw 1.3.0: Multi-Agent Orchestration, Workflow, and Enhanced Knowledge Processing

MateClaw, an open-source AI agent platform, has launched version 1.3.0, focusing on transforming individual AI agents into a cohesive, automated operating system for business processes. The update introduces several key features designed to enhance multi-agent collaboration, knowledge management, and overall system reliability.

Multi-Agent Orchestration and Workflows: The most significant addition is the new Workflow engine, which allows users to compose multiple 'digital employees' (agents) and system actions into linear business processes. This moves MateClaw beyond simple chatbot functionality, enabling complex sequences like enriching customer records, running VIP onboarding, and fanning out communications across multiple channels. Workflows support various step modes, including sequential, fan_out/collect, conditional, await_approval, dispatch_channel, and write_memory. Users can define workflows using a JSON-first DSL with schema validation or leverage natural-language draft generation.

Event-Driven Triggers: To automate workflow initiation, MateClaw 1.3.0 introduces Triggers. These can be configured to start a workflow or send a message to an employee based on various events, such as cron schedules, webhook events, channel_message reception, agent_lifecycle events, content_match, or workflow_completion. The system includes built-in governance features like event deduplication, per-trigger rate limits, bot self-message filtering, and recursion guards to ensure stable operation.

Enhanced Wiki as a Processing Pipeline: The Wiki component has been upgraded from a passive search index to an active processing pipeline. It now supports a transformations engine, allowing users to attach templates to raw materials or pages, run them through an LLM, and save structured output. This enables tasks like contract clause extraction, risk summarization, and KPI distillation. The Wiki can also perform cross-material aggregation (map-reduce) and extract reverse citations, linking synthesis pages back to their exact source chunks.

Improved Agent Tooling and Multimodal Support: Version 1.3.0 addresses previous limitations in tool binding by allowing each digital employee to bind MCP (Multi-Channel Protocol) tools independently, preventing unintended tool access across agents. It also introduces multimodal sidecar routing, where a configured vision model automatically describes images for text-only main models, ensuring appropriate handling of visual input without user intervention. Additionally, the release includes four new JVM-native document rendering tools for generating .docx, .xlsx, .pptx, and .pdf files directly from chat, and enhanced image editing capabilities.

Context Engineering and Reliability: Several under-the-hood improvements in context engineering aim to prevent memory loss during long tasks. The first user message is now anchored, and compaction processes are designed to avoid splitting tool_call/tool_response pairs. Older tool results are preserved as raw data, and spill markers persist across multiple compaction phases. Streaming improvements include sanitizing tool-call arguments to valid JSON in flight and smarter repetition detection.

Additional Updates: The release also features deep tuning for WeCom channels, including approval cards and group chat attribution, an enterprise scenario workbench for tasks like contract review, and deployment enhancements such as optional heavyweight skills and updated Spring AI dependencies. New LLM providers and models, including DashScope-compatible modes, Wanxiang/qwen-image series, Xiaomi MiMo, and Tencent Hunyuan 3D, have been integrated.