Skills

A skill is a capability bundle: instructions plus executors. Each agent has its own enabled skill list, independent of main.

Examples

SkillWhat it enables
read / write / editFiles in the agent workspace
memorySemantic recall across conversations
browse / visionWeb and screenshot understanding
githubRepos, issues, pull requests
gmail / calendarEmail search, summaries, sends, event creation, free/busy checks
home-assistantLights, switches, sensors, scenes, scripts, and thermostat control
ssh-inspectRead-only remote server, service, log, disk, and Docker inspection
agent-sendDelegate to another agent (with allow list)
cronReminders and scheduled jobs

New agents

New agents copy skills from a template agent but drop sensitive defaults until you enable them in the dashboard. Toggle skills per agent from the roster card edit panel.

Intent planning

On each turn, Pasture Protocol classifies intent and selects relevant skills before the LLM loop. Not every enabled skill is loaded every time; only what fits the message.

How skills are selected

User intentTypical skill path
Live informationsearch for summaries, browse for page interaction or proof
Reminder or recurring checkcron stores schedule plus originating chat/channel
Image, screenshot, or webcamvision describes or generates images; browse screenshots can chain into vision
Remember or recallmemory saves explicit notes and searches notes/chat only when the user asks
Project planningproject-workflow links chat to Projects, Missions, Tasks, and progress logs
Home controlhome-assistant resolves friendly names and replies with friendly summaries

Skill boundaries

  • Write-capable integrations should preview important changes and ask for confirmation when the skill requires it.
  • Group chats can deny specific skills even when the assigned agent normally has them.
  • Secrets belong in the Pasture state directory, not in memory notes or project files.
  • Powerful skills should be enabled only on agents that actually need them.

← Agents overview