| name | omp-agents |
| description | Use when a task matches a pre-defined Pi Agent's expertise. Delegate via omp run instead of doing it yourself. Do NOT use for multi-agent orchestration (use team skill). |
High-Frequency Agents
| Agent | Expertise | Typical Tasks |
|---|
researcher | Deep research, web search, information retrieval | "Research latest developments in X", "Compare approaches A vs B" |
reviewer | Skill/Agent quality audit | "Review this SKILL.md", "Check if agent definition is compliant" |
ux-engineer | UI audit, frontend design, style optimization | "Audit UI issues on this page", "Generate Tailwind component" |
oss-researcher | 开源项目代码研究与知识库沉淀 | "pi-mono 如何加载 skill？", "langchain memory 模块怎么设计的？" |
Above is a high-frequency subset. Get the full agent list:
omp list agent
Invocation
omp run <agent-name> --mode stream "task description"
omp run <agent-name> --mode stream -m <model> "task description"
Recommended Models
| Model | When to Use |
|---|
litellm-local/qwen3.5-27b | Default choice, local inference, zero cost |
openai-codex/gpt-5.4 | When higher quality is needed |
Do not use other models.
When to Delegate
Principle: delegate when possible — agents carry specialized skills that outperform raw LLM.
Delegate when:
- Task clearly falls within an agent's expertise
- Task benefits from the agent's specialized skills (e.g. ux-engineer has 18 frontend skills)
- You lack the relevant tools or domain knowledge
Do NOT delegate when:
- Simple task — you can do it faster directly
- Task requires multi-turn interaction —
omp run is one-shot, no follow-up
- Task needs current conversation context — agent cannot see your chat history
Prompt Requirements
- Self-contained — agent has no context from you; include all necessary information in the prompt
- Explicit output format — tell the agent what format you expect (markdown / JSON / code)
- Single responsibility — one task per delegation
