Hiring AI employees, with onboarding, SOPs, and an approval chain

What it does

UniEmployee is a Chinese-enterprise-flavored platform for building and running AI “digital employees” — agents that take on roles like presale support, data analyst, HR, and network operations. Everything is configured through a web UI: persona, model, skills, tools, knowledge bases, SOPs, and connectors. The architecture is a five-layer model (Employee → Workflow/SOP → Skill → Connector → Tool), and a compiler assembles each employee’s config into a runnable agent at startup.

The interesting bit

Skills are written as SKILL.md documents — procedural specs with trigger conditions and steps, seeded into a store the model consults on demand rather than reciting from memory. Critical business flows get hardened into StateGraph state machines with human-approval nodes, so multi-step processes don’t drift. Long conversations auto-compress at ~85% of the context window (or 170k tokens), with full history archived to disk for replay.

Key highlights

• Eight example employees out of the box, covering support, sales, HR, data analysis, and network ops
• Human-in-the-loop: high-risk actions pause the workflow until approved, then resume automatically
• Full observability — every run records LLM/tool inputs, outputs, timing, and token usage, replayable
• Enterprise knowledge ontology: typed entities, relationships, multi-hop queries, source tracing — not just vector chunks
• Persistent per-user, per-employee memory that survives restarts; cron and webhook-triggered tasks reuse the same guardrails
• SQL analysis is fenced in with AST-level read-only validation, table whitelists, and row limits

Caveats

• Editing employee YAML won’t re-seed an existing catalog — a footgun the README itself flags
• Resource-center changes generally require a service restart to recompile cached agents
• External IM channels (WeChat, DingTalk, Feishu) are still on the roadmap; only in-platform web chat works today
• Documentation is primarily Chinese; the English README exists but the ecosystem is clearly domestic-first

Verdict

Worth a look if you’re building agent platforms for real business processes and want a reference for HITL, SOPs, and observability done seriously. Skip it if you wanted a lightweight single-agent library — this is a full platform with a frontend, PostgreSQL, and opinions.