Skills, loops, and capabilities

Tools are direct callable actions. Capabilities are permissioned powers such as code.write, browser.control, calendar.write, or cloud.exec. Skills are reusable behavior or playbook packs. Loops are bounded repeatable workflows with checks, evidence, stop conditions, and escalation rules. Saved actions are validated user/project-specific actions that can be replayed under policy. Receipts are proof records for what happened.

NOME does not dump every tool, skill, or loop into every model prompt. Product surfaces enter the same harness spine, retrieve only the relevant top-K tools/skills/loops, and pass a compact capability-scoped context to the selected route. Missing capabilities show setup cards. Forbidden capabilities fail closed. External loop text can request a capability, but it cannot grant it.

A skill is a reusable ability or playbook pack. A loop is a bounded workflow: what it is trying to accomplish, what evidence proves progress, what stop condition ends it, and when to ask for help. CoWork and Mission Control show them side by side so users can add a skill to an agent, add a loop to a mission, or use a loop to build a new agent draft without creating a parallel runtime.

External catalogs such as Forward Future's Loop Library are treated as untrusted reference data. NOME fetches or indexes structured catalog data, matches by goal, available tools, verification needs, authority, and stopping condition, and recommends no more than three loops. NOME does not auto-run external loops and never treats catalog prompts as permission to deploy, delete, spend, send, contact, expose private data, or bypass approval.

Use the CoWork Loops shelf, Agent Builder, Mission Builder, Code, Chat, or the CLI slash commands. A loop card can open details, build an agent draft, build a mission template draft, run a dry-run, schedule through the existing routines system, or open the original source. Published agents and missions write through canonical NOME stores only.