Symbio provides an AI infrastructure-level framework for multi-agent orchestration, addressing complex engineering challenges beyond basic tool-calling. It targets developers and researchers seeking robust control, observability, and evolvability in agent systems, offering a unified runtime for critical infrastructure modules like scheduling, memory, security, and cost management.

How It Works

Symbio decomposes agent capabilities into distinct, observable, approvable, recoverable, and verifiable infrastructure modules. Its core approach centers on a unified runtime that integrates dynamic DAG scheduling, ontological memory with zero-token reasoning, multi-channel human-in-the-loop (HITL) approvals, and a three-layer prompt injection firewall. This design prioritizes engineering concerns such as task recovery, cost control, and security, enabling agents to operate more reliably and efficiently.

Quick Start & Requirements

• Primary install: pip install symbio
• Initialization: symbio init
• Serve: symbio serve --port 9090
• Prerequisites: Python 3.10+. Development from source requires pip install -e ".[dev,server,ml,otel]".
• Docs: README_en.md

Highlighted Details

• Dynamic DAG Runtime: Supports persistent graph states, event-driven execution, replanning, and breakpoint recovery for complex workflows.
• HITL Multi-Channel Approval: Integrates with WeChat, QQ, Feishu, DingTalk, Telegram, and Slack for approvals, with configurable timeout actions (reject, approve, escalate).
• Ontological Memory: Organizes concepts, entities, and relationships into a queryable, visualizable graph supporting zero-token symbolic reasoning.
• Prompt Injection Firewall: A three-layer defense (input purification, semantic detection, intent auditing) intercepts high-risk inputs before LLM calls, achieving ~65% interception on attack samples.
• Token Cost Optimization: Features five layers of optimization including semantic caching (zero-token responses for similar queries), context pruning, and cost monitoring with budget alerts.
• Data Flywheel: A four-stage closed loop for trajectory capture, failure analysis, SOP distillation, and反哺 optimization, exposed via /api/flywheel/*.

Maintenance & Community

The project is actively developed, with a clear capability ledger differentiating implemented, partially implemented, and planned features. Links to English, Chinese, and Japanese documentation are provided. Community and contribution specifics are not detailed in the README.

Licensing & Compatibility

• License: MIT License.
• Compatibility: Permissive MIT license generally allows for commercial use and integration into closed-source projects.

Limitations & Caveats

Symbio is currently in an Alpha stage. While core functionalities like scheduling, HITL, memory, security, and cost management are implemented, enterprise-grade deployment, robust security isolation, a complete remote Skill ecosystem, the full MCP protocol, a production-ready training backend, and privacy computing features are still under active development or in the planning phase.