Task planning
Turn a vague business goal into dependent sub-tasks, and reorder them as intermediate results arrive.
Home/AI Agent
XNF AgentOS v2.4
Enterprise AI agent
runtime
The value of a large model is not in answering — it is in finishing. AgentOS turns goal decomposition, tool calls, knowledge retrieval and result verification into a running loop, so agents can take real roles inside enterprise systems.
96.8%Intent accuracy in production
99.2%Tool-call success rate
1.4sMedian first-token latency
12+Prebuilt connectors
Architecture
Keeping uncertainty inside guardrails
Agent risk comes from improvisation. We use a fixed execution loop plus strict permission and verification layers, so every step is explainable and reversible.
01 · Perceive
Understand
Parse text, documents, speech and business data; identify the real intent and its constraints.
02 · Plan
Plan
Decompose the goal into executable sub-tasks with dependencies and required tools.
03 · Act
Execute
Call internal APIs within role permissions, read and write data, return structured results.
04 · Verify
Verify
Self-check against policies and knowledge; re-plan or hand over to a human when below bar.
IngressInteraction layer
Web portalWeComDingTalkMini appVoiceOpen API
OrchestrationCore
PlannerRouterReflectionMulti-agentHuman-in-the-loop
CapabilityTools & knowledge
RAGTool useWorkflow engineMultimodalExtraction
ModelModels
Private LLMCloud APIEmbeddingVector DBASR / TTS
GovernanceTrust
Access controlAuditEvaluationCostCompliance
Agent Capability
Six modules that decide whether an agent can ship
Tool use
Interfaces are described by schema; the agent picks tools and arguments, with retries and graceful degradation.
Memory
Short-term session memory plus long-term user and business memory keeps context across turns.
Multi-agent collaboration
An orchestrator dispatches to specialist agents and arbitrates when their conclusions conflict.
Evaluation & regression
Scenario test sets track accuracy, completion and hallucination rates across releases, so upgrades stay safe.
Full-chain observability
Complete traces of reasoning, calls and outputs support replay, root-cause analysis and cost attribution.
Live Trace
Watch a real agent trace
Pick a scenario and watch the agent decompose the task, call tools, self-check and deliver an actionable result. Sample data is anonymised.
Select Scenario
Agent Runtime · Trace
Production metrics
Anonymised statistics from live deployments in H1 2026.
Tech Stack
An open stack, no vendor lock-in
Built for private and domestic-cloud environments: models can be swapped, tools extended.
Function callingMCPGraph-style orchestrationHybrid RAG
BGE / RerankMilvusPostgreSQLRedis
ElasticsearchKubernetesvLLMOllama
DockerOpenTelemetryWeCom / DingTalk APIs
◆
Model-agnostic by design. The same agent logic runs on cloud APIs or locally hosted models, so cost and compliance can drive the choice per scenario.
Trust & Safety
The three things enterprises ask about, designed in
Data stays in your network
Full on-premise deployment keeps documents, vectors and inference inside the enterprise, with field-level masking and isolation.
Permissions and human sign-off
Tool calls inherit role permissions; payments, refunds and outbound messages require explicit confirmation and are logged.
Measurable outcomes
Business test sets validate before launch; accuracy and handover rates are monitored afterwards, so scope expands on evidence.
Next Step
Take one real scenario and ship your first agent
We cover scenario assessment, prototype validation and production deployment — usually a usable first version within 2–4 weeks.