| name | om-create-ai-agent |
| description | Build or extend a typed standalone module AI agent/tool, approval-gated mutation, UI part, orchestrator file agent/subagent, attachment/artifact flow, or AI override. Use for "create AI agent", "add AI tool", "agent orchestrator", "subagent", "AI workflow", or "stwórz agenta AI". |
Create a Safe AI Agent
Choose module AI versus file-agent orchestration, then implement typed inputs/outputs, authorization, scope, storage, and validation.
Workflow
- Read
.ai/guides/ai-workflows.md and choose typed module agent/tool, low-level MCP/OpenCode/Code Mode, orchestrator, or workflow with references/surface-selector.md.
- For module agents/tools, follow
references/module-agents-and-tools.md: discovery files, definitions, model factory, tools, ACL/setup, approval mutations, UI parts, loop budgets, and generation.
- For file-agent/orchestrator/subagent work, invoke
om-framework-context for the installed orchestrator module, then follow references/orchestrator-agents.md for outcomes, samples, embedded skills, bounded delegation, outputs, and resume/error states.
- Follow
references/attachments-and-overrides.md for authorized attachments/artifacts, cleanup, extensions, replacements, and disable behavior.
- Test missing provider, denied ACL/scope, tool validation, budgets, approval/cancel/expire/stale version, artifact authorization, and generated registration.
An authorized AI file upload stays on the attachment branch; add integration or data-model context only for a custom transport, storage provider, or app-owned persistence.
An agent/tool that scores or explains customer, contact, lead, or deal records must load .ai/guides/modules/customers/index.md; read-only access still targets that host.
Rules
- Every data tool is scoped, feature-gated, schema-validated, bounded, and serializable.
- Every mutation is marked and routed through ; approval precedes the command write.
