Agent · AI Programming

XXL-AI v1.4.0 Ships Human-in-the-Loop Components, Agent Trace, and a Git Tool

By 许雪里 · Nov 1, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

For teams running agents inside their own infrastructure, these changes bring the two missing pieces of production agent work: a sanctioned path for human approval on tool actions, and a trace view that turns agent failures into inspectable runs. The unified interaction components mean you don't have to hand-roll your own approval UI, and the fingerprint-based cache fixes stop model-switch staleness.

Summary

XXL-AI v1.4.0 focuses on making agents more controllable and debuggable. The desktop client now supports a unified human-machine interaction flow: when an agent needs a decision, it sends an interaction request, and the human confirms through standard components for permission approval, multi-choice selection, or free-text input. A developer mode exposes trace-level observability, showing session conversations, tool calls, model API interactions, and memory operations for each run. The release also adds a Git tool for project-space review and push, a SKILL marketplace that installs from GitHub, and agent-level customization including instruction presets and memory toggles.

The cloud version gets security and consistency fixes. Homepage report data is now tenant-isolated to prevent cross-space leaks. The Milvus vector store's cache key includes a model fingerprint, so changing models invalidates stale entries. Agent model objects refresh every five minutes and update actively when config fingerprints change. Both versions receive the same Markdown and thinking-rendering optimizations: better fonts, code highlighting, and capped scrollable thinking panes.

Takeaways

— The desktop client adds a unified human-machine interaction flow where agents request user confirmation during task execution.

— Standard interaction components cover permission confirmation, multi-option selection, and user input.

— Developer mode enables agent observability via Trace, showing session, tool calls, model API interactions, and memory operations.

— A new Git tool supports project-space change review and push.

— SKILLs can be installed from GitHub, with multi-SKILL detection, caching, and expiry cleanup.

— Agents now support custom instructions, memory toggles, and memory management.

— Preset agent instructions are provided for quick setup.

— The terminal, browser, and file components move to a unified right-side panel.

— The cloud version adds tenant isolation for report data.

— MilvusVectorStore cache keys now include model fingerprint to avoid stale caches.

— Model objects refresh every 5 minutes and actively update on config fingerprint changes.

Conclusions

The standardized human-in-the-loop components imply that agent autonomy on local systems is being designed with a mandatory approval gate, not just an optional safety feature.

Adding trace observability as a developer mode suggests that debugging agent behavior is still a specialized activity, and the trace covers more than model calls, including tool calls and memory operations.

The cloud version's 5-minute model refresh with fingerprint invalidation is a pragmatic answer to a common staleness problem, especially when model configs change frequently.

Concepts & terms

MCP

Model Context Protocol: an open protocol for connecting AI models to external tools and data sources, letting agents call tools like web scraping or file operations.

SKILL

In XXL-AI, a packaged, reusable capability that an agent can invoke, installable from a marketplace or from GitHub; typically bundles instructions and tool definitions.

RAG

Retrieval-Augmented Generation: a technique where an AI model retrieves relevant documents or data before generating a response, improving accuracy with external knowledge.

Agent Trace

Observability output that records each step an agent takes, including session chats, tool calls, model API interactions, and memory operations, used for debugging.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗