How OpenAkita Makes Three AIs Collaborate Automatically

Multi‑Agent Collaboration Core

OpenAkita implements a multi‑Agent workflow. A user request such as “create a competitive‑analysis report” is split into three concurrent Agents: a Search Agent gathers data, an Analysis Agent processes the data, and a Writing Agent composes the final report. The AgentOrchestrator schedules these tasks, while the AgentInstancePool limits delegation depth to five layers to avoid runaway recursion. Each Agent’s status is visualized on a neural‑network dashboard.

When an Agent stalls, the system automatically switches to an alternative strategy instead of failing. The underlying reasoning follows the ReAct loop – think → act → observe – with explicit checkpointing that enables rollback on errors.

System Architecture

The architecture consists of six layers:

Tauri + React desktop application (user interface)

Inference engine (model execution)

Agent scheduling layer (Orchestrator, InstancePool)

Memory subsystem (working, core, dynamic retrieval)

Tool layer (89 built‑in tools)

Integration layer for six IM channels

Each layer has a clearly defined responsibility, avoiding monolithic code.

Built‑in Tools

OpenAkita ships with 89 tools organized into 16 categories. Example scenarios:

Schedule a daily 9 am task that collects competitor updates and posts to a DingTalk group.

Send a voice message in Telegram; the system transcribes and executes the command.

Control a local Excel instance to fill data directly on the desktop.

Plan Mode automatically decomposes complex tasks into sequential steps, displays a progress bar for each step, and rolls back to an alternative plan if a step fails.

Zero‑Command‑Line Installation

Installation is performed via a graphical installer: download the package, double‑click, run the wizard, enter the API key, and start the application. The Python runtime is isolated from the host system, and no manual git clone or YAML editing is required.

The desktop UI provides eleven panels for full graphical control.

Model and IM Support

Model providers are not hard‑wired; the system supports over 30 providers (e.g., DeepSeek, Tongyi Qianwen, Kimi, Claude, GPT, Gemini). If a model endpoint fails, OpenAkita automatically falls back to a backup model.

Six IM platforms are integrated:

Feishu – WebSocket/Webhook with card messages and event subscription.

WeCom – Smart Robot with streaming replies and proactive pushes.

DingTalk – Stream WebSocket, no public IP required.

Telegram – Webhook/Long Polling, markdown output, proxy support.

QQ – WebSocket/Webhook supporting group, private, and channel chats.

OneBot – WebSocket compatible with NapCat/Lagrange.

Agents can be invoked via @‑mentions or voice commands within these chat tools.

Three‑Layer Memory System

Memory is divided into:

Working memory – current task context.

Core memory – user profile and preferences.

Dynamic retrieval – historical experiences.

The system supports seven memory types and retains information across sessions (e.g., remembering a user’s preferred writing style).

Eight preset personas are available (default, technical expert, “boyfriend” mode, “girlfriend” mode, Jarvis, household assistant, business, family) and can be switched with a single command. After three hours of inactivity the assistant proactively checks in.

Data Security and Self‑Evolution

All data is stored locally; no cloud upload occurs. A POLICIES.yaml file defines safety policies, requiring user confirmation for risky actions and enforcing resource budgets. Runtime monitoring detects dead loops.

Each night at 04:00 AM the system runs an AI‑driven self‑diagnosis: it parses error logs, attempts automatic fixes, and if a missing skill is detected, searches GitHub or generates the skill on the fly.

Current Release

Version v1.25.0 with more than 758 commits. Installers are provided for Windows, macOS, and Linux at: https://github.com/openakita/openakita/releases Python package installation: pip install openakita[all] Source repository: https://github.com/openakita/openakita

Community channels: Discord (https://discord.gg/vFwxNVNH) and X (formerly Twitter) @openakita.

Evaluation Insight

Multi‑Agent collaboration is a research focus of major AI companies. OpenAkita demonstrates that an open‑source stack can provide robust agent orchestration while lowering the entry barrier for non‑programmers.