# Platform Agents

> Build or repair an AgenticOS agent: choose capabilities and MCP tools, write its prompt, bind skills and knowledge, test failure cases and publish a verified version. Use for agent setup and configuration, not a one-off specialist.

- Skill: `vstorm-co/platform-agents` (Agent Skill)
- Install (CLI): `npx get-skill add vstorm-co/platform-agents`
- Raw SKILL.md: https://api.skillmd.com/api/skills/vstorm-co/platform-agents/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: vstorm-co (https://skillmd.com/u/vstorm-co)
- Updated: 2026-11-12
- Page: https://skillmd.com/skills/vstorm-co/platform-agents

---


# Design an agent around an observable job

Use for consequential agent design or diagnosis, not a routine edit of one setting. MCP discovery and schemas define available capabilities, model configuration and draft/publish operations.

Settle the deliverable, audience, authoritative sources, permissible effects and evidence of completion. Inspect the existing agent before proposing a new one. Start with the capabilities required for that job; another service's account, a local computer and the platform's own records are separate access boundaries.

Write instructions as a working contract: role, source precedence, decision policy, failure handling and completion criteria. Explain the judgments that tool descriptions cannot supply. Avoid copying an API manual, enumerating every tool, or requiring ceremonial setup calls. Keep stable facts in Context, reusable methods in skills, preferences in Memory and pending work in Tasks.

Delegate bounded investigation and review with relevant context and acceptance criteria. Keep integration with the lead agent. Choose a workflow for a deterministic pipeline, an agent for judgment and a routine for recurrence; use the fewest mechanisms that meet the goal.

Test the draft on an ordinary request, missing evidence, refused access and a hostile document. Check actual effects and cost. Draft testing may call real services; keep its effects within the person's authorization. Publishing and enabling unattended work are distinct decisions. Return the resource, checks performed, remaining blockers and whether it is actually published.

