Skip to content
Open app

Platform agent

The platform agent is a conversational way for governing teams to inspect their skill estate — coverage, recurring failures, eval health — and draft improvements, without stitching together REST calls. Ask it “are we covered?”, “what keeps breaking?”, or “what should we require?”, and it answers by calling the same governance data the API exposes.

The agent is a stateful chat agent (Cloudflare Agents SDK) rather than a REST endpoint — it runs over a WebSocket. There is one instance per organization, reached at /agents/skillist-agent/{orgId} on the API. The instance name is the org id, so every answer is scoped to that org and cannot read across orgs. It never invents skills or numbers: each answer comes from a tool call.

The agent has six governance tools. Five are read-only; only draft_improvement writes, and even then only as pending feedback a human must approve.

ToolWhat it doesWrites?
get_coverageRequired-skill coverage across published, covered, and activated — plus driftNo
list_recurring_failuresOpen and drafted recurring failure patterns, most frequent firstNo
list_required_skillsThe org's declared required-skill baselineNo
recommend_required_skillsPopular registry skills the org does not yet requireNo
flag_stale_evalsSkills whose latest eval failed, is over ~30 days old, or never ranNo
draft_improvementOpens agent-sourced pending feedback against a skill for human approvalYes (pending only)

draft_improvement never edits a skill directly. It lands in the review inbox as pending feedback — approving it triggers the same AI-draft improvement flow described in Self-improving skills. The verified user who prompted the chat is attributed on the drafted item.

  • Coverage — “Are we covered? What’s drifting?” → get_coverage
  • Failures — “What keeps breaking? Show the top failures.” → list_recurring_failures
  • Baseline — “What do we require? What are we missing?” → list_required_skills, recommend_required_skills
  • Eval health — “Which skills need re-evaluation?” → flag_stale_evals
  • Act on it — “Draft a fix for pdf-tools about the missing interpreter.” → draft_improvement