HalosHalosDocs

Skills & Trust

Skills determine what your agents can do. Trust tiers determine what they're allowed to do. Together they form a progressive permission system where agents earn capabilities over time.

Skills

Skills are modular capabilities that you install on individual agents. They come from ClawHub, the OpenClaw skill marketplace, and from the built-in Halos skill library.

How Skills Work

Each agent starts with a set of core skills that every agent gets regardless of type:

| Core Skill | What It Does | |-----------|-------------| | Web Browsing | Navigate websites, interact with pages, extract data via Stagehand + Browserbase | | Web Search | Search the internet for real-time information via Brave Search | | File Management | Handle file uploads, attachments, and storage | | Shell Commands | Execute commands for automation tasks |

On top of these, each agent type gets specialty skills based on its domain:

| Agent Type | Default Specialty Skills | |-----------|------------------------| | Email | Email Triage, Draft Replies, Smart Labels | | Calendar | Schedule Management, Conflict Detection, Smart Reminders | | Research | Web Search (Deep), Article Summary, Trend Tracking | | Code | Code Review, Debugging, Code Generation | | Photo | Core skills only (customized via system prompt) | | Custom | Core skills only (you curate from scratch) |

Installing Skills

From the Skills page or Skill Library:

  1. Browse available skills by category or search by name
  2. Click a skill to see its description, example prompts, and requirements
  3. Choose which agent to install it on
  4. The skill is available immediately — hot-installed, no restart needed

Skill Recommendations by Use Case

Here are some useful skill combinations beyond the defaults:

Personal assistant agent:

  • Web Search + Article Summary + Note Taking
  • Good for: research, learning, staying informed

Project manager agent:

  • Task Management + Status Reports + Meeting Notes
  • Good for: tracking work, generating updates

Writer/content agent:

  • Content Generation + SEO Analysis + Grammar Check
  • Good for: blog posts, social media, documentation

Data analyst agent:

  • Data Analysis + Visualization + Report Generation
  • Good for: spreadsheets, dashboards, insights

Skill XP

Each skill has its own XP and level, separate from the agent's overall level. Skills level up as the agent uses them:

  • Level 1 — Basic capability
  • Level 5 — Proficient
  • Level 10 — Expert

Higher skill levels indicate more experience with that capability.

ClawHub

ClawHub is the OpenClaw skill marketplace — a community-maintained catalog of agent skills. Skills are defined as portable Markdown files (SKILL.md) with structured YAML frontmatter that describes:

  • What the skill does
  • What tools and integrations it requires
  • Example prompts
  • Difficulty level (beginner, intermediate, advanced)
  • Categories and tags

You can browse ClawHub skills from the Skill Library page in Halos, which shows available skills with filtering by category, search, and difficulty.

Native Tools

Beyond installable skills, every agent has access to a set of native tools — built-in capabilities powered by TypeScript functions that call APIs directly. These tools are always available when their requirements are met.

Always Available

| Tool | What It Does | |------|-------------| | knowledge_search | Search your personal knowledge base using Pinecone vector similarity | | knowledge_store | Save a new piece of knowledge (fact, preference, note) | | manage_todos | List, create, complete, and update todos |

When Integration Connected

| Tool | Requires | What It Does | |------|---------|-------------| | read_emails | Gmail (Google OAuth) | Read your Gmail inbox, search by query | | read_calendar | Google Calendar | View upcoming events | | create_event | Google Calendar or Apple Calendar | Create new calendar events | | notion_read | Notion OAuth | List pages or read page content | | notion_write | Notion OAuth | Create pages or append content | | send_email | Agent email configured | Send emails from the agent's address | | read_inbox | Agent email configured | Read the agent's email inbox |

When API Key Configured

| Tool | Requires | What It Does | |------|---------|-------------| | web_search | Brave API key | Search the internet | | web_scrape | Firecrawl API key | Extract markdown content from any URL | | browser_navigate | Browserbase | Open any URL in a cloud browser | | browser_act | Browserbase | Click, fill, scroll using natural language | | browser_extract | Browserbase | Pull structured data from pages | | browser_observe | Browserbase | See what's on the current page |

Head Agent Only

| Tool | What It Does | |------|-------------| | manage_agents | Create, update, and manage specialist agents (head agent only) |

Tools are composed dynamically per-request. If you haven't connected Google, the email/calendar tools simply don't appear — no "tool not available" errors, just clean scoping.

Trust Tiers

Trust tiers are a progressive permission system. Every agent starts at the lowest tier and can be promoted as you build confidence in its behavior.

Why Progressive Trust?

You wouldn't give a new employee full admin access on day one. The same principle applies to AI agents. Progressive trust means:

  • You stay in control while learning what an agent can do
  • Agents prove themselves before getting more power
  • Mistakes at lower trust levels have minimal consequences
  • You can always demote an agent if it's not performing well

The Four Tiers

Observer (Default)

The starting tier for all new agents.

  • Can: Read and observe data, summarize, analyze, suggest actions
  • Cannot: Modify, send, delete, or create anything
  • Philosophy: Look but don't touch

This is intentionally restrictive. A new agent should prove it understands your needs before getting more power.

In practice: Your email agent can read and summarize your inbox, flag urgent items, and draft replies for your review — but it can't send anything.

Helper

The first promotion. The agent can assist more actively.

  • Can: Read data, draft content, organize, label, categorize
  • Cannot: Send, publish, or delete without your explicit approval
  • Philosophy: Prepare but don't execute

In practice: Your email agent can draft replies, organize your inbox with labels, and sort by priority — but you still approve every outgoing email.

Partner

A significant trust upgrade. The agent acts autonomously on routine tasks.

  • Can: Act autonomously on low-risk, routine tasks. Send pre-approved types of responses
  • Cannot: Message new contacts, take financial actions, or delete without approval
  • Philosophy: Handle the routine, escalate the unusual

In practice: Your calendar agent automatically accepts recurring meeting invites and sends reminders, but asks before scheduling with a new contact or canceling an important meeting.

Trusted

The highest tier. Near-full autonomy.

  • Can: Act autonomously on most tasks
  • Cannot: Bulk deletions, purchases, or external sharing without approval
  • Philosophy: Full trust with safety rails

In practice: Your email agent sends routine replies, archives low-priority threads, and only escalates genuinely unusual situations. Even at this level, destructive or high-stakes actions still require your sign-off.

Permissions

Within each trust tier, individual permissions provide granular control.

How Permissions Work

Each agent type has specific permissions that unlock at certain levels:

Email Agent Example:

| Permission | Unlock Level | Description | |-----------|-------------|-------------| | Read data | Level 1 | Access and read connections | | Summarize | Level 1 | Create summaries and digests | | Suggest actions | Level 1 | Recommend actions for approval | | Draft replies | Level 3 | Write reply drafts for review | | Apply labels | Level 5 | Organize emails with labels | | Send emails | Level 8 | Send emails on your behalf | | Delete emails | Level 15 | Delete or archive emails |

Enabling Permissions

Permissions use a two-layer system: level-gating + manual toggle.

  1. Locked (grayed out) — Agent hasn't reached the required level yet
  2. Unlocked, disabled — Available but not turned on
  3. Unlocked, enabled — Active and the agent can use it

Even when a permission unlocks, it starts off by default. You decide what to enable. This means you're never surprised by an agent suddenly gaining new powers.

Agent Leveling

Agents gain XP by completing tasks and earn levels over time.

How XP Works

  • Completing a task: +10 XP
  • XP required scales with level: 100 + (level x 50) XP per level
  • Level 1 needs 100 XP, Level 2 needs 150 XP, Level 3 needs 200 XP, etc.

What Leveling Unlocks

  • New permissions become available (see above)
  • The agent's profile shows its level and XP progress bar
  • Higher levels signal a more experienced, trusted agent

The Heartbeat

Each agent has a configurable heartbeat — a proactive check-in interval. This is what makes your agents proactive rather than just reactive.

Heartbeat Intervals

Set the interval on the agent's profile:

| Interval | Best For | |----------|---------| | Disabled | Agents you only use on-demand | | Every 15 minutes | High-priority monitoring (urgent email triage) | | Every hour | Balanced proactive updates (default) | | Every 8 hours | Twice-daily summaries | | Once a day | Daily digest style |

What Heartbeats Do

During a heartbeat, the agent proactively:

  • Checks connected integrations for new data
  • Summarizes anything notable since the last check-in
  • Suggests actions you might want to take
  • Reports on completed background tasks

Example heartbeat from an email agent:

"3 new emails since your last check. One from your manager about the budget review — looks time-sensitive. Two newsletters that can wait. Want me to draft a reply to your manager?"

When to Use Heartbeats

  • Email agents — Set to hourly or more frequent so you catch urgent items early
  • Research agents — Set to daily for industry news digests
  • Calendar agents — Set to daily for next-day schedule prep
  • Code agents — Usually disabled, since you interact with them on-demand

Stats & Tracking

Every agent tracks its own performance:

  • Tasks completed — Total tasks the agent has handled
  • Time saved — Estimated minutes/hours saved for you
  • Accuracy — How often the agent gets it right

These stats appear on the agent's profile page and contribute to the platform-wide time-saved tracker visible in your settings.

Next Steps