Support DeepSeek, Qwen, Zhipu, Mock and other LLM providers with agent-level provider and model configuration.

Different agents can use different LLMs, selecting the most suitable model based on task characteristics to balance effectiveness and cost.

1)    Agent-level provider configuration: each agent independently configures LLM provider, model name, temperature parameters.

2)    Flexible switching: switch agent's LLM provider anytime without redeployment, effective immediately.

3)    Mock mode: use Mock LLM during development and testing without consuming real tokens, reducing development costs.

4)    Provider health check: auto-detect LLM provider availability, auto-failover to backup provider on failure.

AI agents can call ERP system tools and APIs to execute actual business operations, the essential difference from chatbots.

Through Target Model plus JSON Schema auto-generating tool_call, AI can create orders, query inventory, generate reports, and approve workflows.

1)    Each agent can bind multiple ERP tools, tool list managed through configuration without code changes.

2)    JSON Schema auto-generates tool_call: tool inputs/outputs defined via JSON Schema, LLM auto-generates correct call format.

3)    Multi-step reasoning: agents can execute multi-step reasoning, querying data then analyzing then executing operations.

4)    Tool result feedback: tool execution results auto-fed back to LLM, which decides next action based on results.

5)    Tool permission control: different agents can use different tool sets, implementing least privilege principle.

Multiple AI agents can collaborate to complete complex tasks that a single agent cannot handle alone.

Sales AI can call Inventory AI to check availability, call Finance AI to calculate discounted prices, call Production AI to check capacity, enabling end-to-end business collaboration.

1)    Inter-agent communication: supports RPC, message queue, and MCP protocols for different scenarios.

2)    Collaboration orchestration: orchestrate multiple agents' collaboration flow with defined execution order and dependencies.

3)    Context passing: agents can pass context information like customer info and order data, avoiding redundant queries.

4)    Collaboration audit: complete audit logs for inter-agent collaboration, traceable inputs and outputs for each agent.

5)    Load balancing: auto-distribute load during multi-agent collaboration, preventing single agent overload.

After user input, the system auto-identifies intent and routes to the most appropriate agent and workflow.

Not simple keyword matching, but multi-layer routing based on regex matching plus confidence scoring plus workflow binding.

1)    5 routing types: exact match, fuzzy match, semantic match, workflow binding, dynamic reasoning routing.

2)    Regex matching plus confidence scoring: each routing rule has regex pattern and confidence threshold.

3)    Workflow binding: route match can bind to specific workflow for structured business process handling.

4)    Dynamic reasoning routing: LLM dynamically selects the most appropriate processing path based on user intent.

5)    Routing priority: multiple routing rules can set priority, matched in priority order.

Agents support chat, workflow, and dual execution modes for different business scenarios.

The same agent can serve as a chat assistant answering questions and execute structured workflows for business operations.

1)    Chat mode: free conversation mode for consulting, Q&A, guidance and other unstructured interactions.

2)    Workflow mode: structured flow execution for approval, order processing and other structured scenarios.

3)    Dual mode: agent supports both chat and workflow, auto-selecting based on user intent.

4)    Mode switching: agents can switch execution mode at runtime without redeployment.

Every AI call has audit logs, cost budgets, and data masking, with full AI product lifecycle management.

From AI agent creation to retirement, the entire lifecycle is governable, traceable, and auditable.

1)    Audit logs: record token consumption, execution time, tools called, and results for every AI call.

2)    Cost budgets: budgets by agent, project, department with overflow alerts for fine-grained cost control.

3)    Data masking: auto-mask sensitive data before AI calls, audit logs don't leak sensitive information.

4)    AI product lifecycle management: version, baseline, change, release, audit, complete PLM system.