UIP 智能体

Scaffold, validate, refresh, review, and manage low-code agents using the `uip agent` tool.

uip agent is the Agent tool: it scaffolds, validates, migrates, and reviews low-code agents — the agent.json-driven projects authored in Agent Builder / Studio Web. Install it with uip tools install agent, or invoke uip agent … and let auto-install run. See Tools (plugins) for the install model and uip tools for the command that manages tool installation.

This tool is distinct from the coded agent tool (uip codedagent), which targets Python agents built with LangGraph / LlamaIndex / OpenAI Agents. uip agent only operates on low-code agent projects: a directory containing agent.json, entry-points.json, project.uiproj, an evals/ tree, and features/ / resources/ folders — or, for an agent embedded in a Maestro Flow, a UUID-named subdirectory inside the flow project (--inline-in-flow).

uip agent has no pack, publish, deploy, push, pull, run, list, share, file, config, context, escalation, input, or output command — packaging and shipping an agent project goes through uip solution instead (the agent project is one project inside a solution). Resource authoring (tools, contexts, escalations, I/O schema) is done by hand-editing agent.json, not through a dedicated CLI verb.

什么是 UiPath 智能体

低代码智能体由单个agent.json文件以声明方式进行描述 — 模型、提示、输入/输出架构、工具、上下文、升级和设置。支持文件描述了 Studio Web 用于调用智能体、默认评估集以及任何 RAG/HITL 资源的入口点。

The real lifecycle covered by uip agent is:

1. Scaffold a project on disk with uip agent init (standalone, or --inline-in-flow inside a Maestro Flow project). By default the project auto-registers into a surrounding solution's .uipx, or a new <Name>Solution is scaffolded automatically when none exists.
2. Author by hand-editing agent.json (and the files under resources/, evals/).
3. Refresh the project (uip agent refresh) after every edit — applies pending schema migrations and regenerates entry-points.json / bindings_v2.json. This is the only command that writes derived files.
4. Validate the project (uip agent validate) — strict, read-only: answers "would Studio Web accept this directory as-is?" without writing anything.
5. Review the project against a deterministic rule set (uip agent review) for a scored quality verdict, and optionally record the grade over time (uip agent review-history add).
6. Configure LLM models (uip agent model list), memory spaces (uip agent memory), guardrails (uip agent guardrails), and tools (uip agent tool list) as needed.
7. Debug the agent against Studio Web (uip agent debug) to run it end-to-end with sample inputs before shipping.
8. Ship with uip solution pack / publish / upload / deploy — uip agent has no packaging or deployment verb of its own.
9. Evaluate the agent against an evaluation set (uip agent eval run start) — works against a cloud-pushed solution (--solution-id, or auto-detected from SolutionStorage.json) or, when neither is available, an implicit local-workspace evaluation path.

动词

命令	用途
初始化	在磁盘上构建新的智能体项目（独立或内联在流程项目中）。
验证	Strict read-only check: would Studio Web accept this project as-is?
refresh	Apply pending schema migrations and regenerate entry-points.json / bindings_v2.json.
review	Score the project against a deterministic rule set (verdict, score, grade, issues).
review-history	Record a review grade in the project's local review-history file.
tool list	List the tool resources configured on the agent.
memory	Attach, list, and remove memory spaces and memory items.
评估	管理评估程序、评估集、测试用例和评估运行。
guardrails	List guardrail definitions, BYOG guardrails, LLM-as-judge models, and the guardrail catalog.
model	List LLM models available for the authenticated tenant.
调试	Run the agent against Studio Web with sample inputs and stream the result.

大纲

uip agent <verb> [subverb] [arguments] [options]

始终使用uip agent ，从不使用uip agents或uip a 。agent是工具的commandPrefix （请参阅uip tools list ）。

所有uip agent动词都遵循四个全局选项（ --output 、 --output-filter 、 --log-level和--log-file ）。退出代码遵循共享合同。

端到端流程

# 1. Scaffold (auto-registers into the surrounding solution, or creates one)
uip agent init ./my-agent --model gpt-5.4

# 2. Refresh after any edit to agent.json (migrates + regenerates derived files)
uip agent refresh ./my-agent

# 3. Validate the project is ready for upload (strict, read-only)
uip agent validate ./my-agent

# 4. Review for quality issues before shipping
uip agent review ./my-agent

# 5. Ship via the solution the project was registered into
uip solution pack . ./dist
uip solution upload .

# 6. Evaluate (against the cloud project once pushed, or the local workspace)
uip agent eval run start --set default --path ./my-agent --wait

# 7. Debug a specific run against Studio Web
uip agent debug ./my-agent -i '{"input":"hello"}'

身份验证

debug, eval run *, guardrails llm-as-judge-models, and model list call Studio Web / the LLM gateway and require an active CLI session. Run uip login first. See Authentication for the session model.

init, validate, refresh, review, review-history, tool list, memory, guardrails list, and guardrails catalog are local-only and do not require a session.

另请参阅

• 概念：工具（插件） — 插件模型和版本固定。
• uip solution — packaging, publishing, and deploying the solution an agent project lives in.
• 身份验证— 会话、租户和文件夹。
• 全局选项，退出代码。