Agents

Agent Skills

An agent skill is a bounded workflow that can request memory, perform a task, and optionally save the outcome. Good skills are narrow, testable, and explicit about what they remember.

Skill Template

type NeutronSkill = {
  id: string;
  name: string;
  purpose: string;
  allowedScopeTypes: string[];
  canReadMemory: boolean;
  canWriteMemory: boolean;
  canDeleteMemory: boolean;
};

Example: Support Reply Skill

const support_reply_skill = {
  id: "support_reply",
  name: "Support reply",
  purpose: "Draft a customer support response using account preferences and recent outcomes.",
  allowedScopeTypes: ["account", "workspace"],
  canReadMemory: true,
  canWriteMemory: true,
  canDeleteMemory: false
};

Prompt Guidance

Use Neutron memory as scoped supporting context. Prefer the current customer message and live product data when they conflict with saved memory. Do not reveal private memory labels, hidden IDs, or sensitive details unless the current user is authorized and the detail is necessary.

Write Policy

Save memory only when it is likely to help future work:

  • stable preferences
  • durable facts
  • resolved decisions
  • reusable summaries
  • important task outcomes

Do not store secrets, provider keys, payment details, raw credentials, or unnecessary personal data.