Skip to main content

Skills — Reusable AI Capabilities

A Skill is a versioned, reusable capability that an agent can invoke. Skills encapsulate prompts, connector calls, business rules, or output guardrails — keeping your agents clean and composable.

Scenario: A marketing team creates a "Tone Rewriter" skill that rephrases any text to match their brand voice. The same skill is reused by the Blog Writer agent, the Social Post agent, and the Email Copywriter agent — with consistent, tested prompting.

Skill types

Type What it does
llm Calls the language model with a prompt template. Most common type.
connector Fetches or sends data to an external service (API, database, webhook).
rule Evaluates deterministic business logic — no LLM call. Fast and cheap.
hybrid Combines LLM + rules/connectors in a single skill invocation.
guardrail Validates or blocks output that doesn't meet quality/safety criteria.
multi_model_consensus Calls multiple LLM providers and merges or majority-votes the results for higher accuracy.
code_execution Runs generated code in a sandboxed environment and returns the output.
browser Controls a headless browser to scrape pages, fill forms, or interact with web UIs.
runpod_endpoint Calls a serverless RunPod endpoint for GPU-accelerated inference or custom model serving.
runpod_pod Spins up a persistent RunPod pod for long-running GPU workloads.
gpu_compute Generic GPU compute skill — fine-tuning, embedding generation, or batch inference.

Versioning

Every time you save a skill, FleetQ creates a new SkillVersion. The active version is used by all agents referencing this skill. You can:

  • View all versions from the skill detail page
  • Roll back to any previous version
  • Track what changed between versions (diff view)
Always test a new skill version on a non-production experiment before rolling it out to all agents. Use the Execute Skill button on the skill detail page for a quick sanity check.

Risk levels

Each skill has a risk level that determines whether human approval is required before execution:

Low

Auto-execute. No approval needed.

Medium

Configurable — can require approval.

High

Always requires human approval before execution.

Critical

Requires approval + owner confirmation. Reserved for irreversible actions.

Cost estimation

FleetQ's SkillCostCalculator estimates token usage before a skill runs, reserves the budget with a 1.5× safety multiplier, then settles the actual cost after completion. You're never surprised by overspend.

JSON schema validation

Define input and output schemas for any skill. FleetQ validates data against the schema at runtime, rejecting malformed inputs before they reach the LLM and ensuring outputs conform to your expected structure.

Example output schema 
{
  "type": "object",
  "required": ["summary", "score"],
  "properties": {
    "summary": { "type": "string", "maxLength": 500 },
    "score":   { "type": "number", "minimum": 0, "maximum": 100 },
    "tags":    { "type": "array", "items": { "type": "string" } }
  }
}