Core Concepts

Capabilities

Capabilities are reusable skills that agents can use to accomplish specific types of work. Think of them as tools in a toolbox — an agent picks up the capabilities it needs for each task.

What is a capability?

A capability is a set of instructions that teaches an agent how to do something specific. Examples:

  • PDF export — generate PDF documents from content
  • Presentation builder — create slide decks (PPTX)
  • Code review — systematic code review with specific criteria
  • Web research — search the web and synthesize findings
  • Internal communications — draft professional emails and memos

Capabilities are defined as SKILL.md files — markdown documents with instructions, examples, and procedures. (The on-disk filename stays SKILL.md for compatibility with the underlying runtime; in Exponential we surface them as capabilities.)

The three tiers

Capabilities follow a trust hierarchy. Higher-trust capabilities are resolved first.

Certified capabilities

Verified and maintained by the Exponential team. These are pre-installed and auto-resolve — no confirmation needed. They cover common, high-value skills.

Examples: pdf, pptx, xlsx, docx, internal-comms.

Local capabilities

Capabilities you create and store on your machine. They're discovered automatically from:

  • ~/.claude/skills/{capability-name}/SKILL.md — available globally
  • your-project/.claude/skills/{capability-name}/SKILL.md — available in that project only

Local capabilities resolve after certified capabilities but before community capabilities.

Community capabilities

Capabilities shared by other users through repositories. These require user confirmation before installation — you review what the capability does before granting access.

Community capabilities are the lowest trust tier but offer the broadest coverage.

How capability resolution works

When an agent plans a task, Exponential automatically figures out which capabilities are needed:

  1. Agent proposes a plan — the plan describes what needs to happen.
  2. System detects requirements — based on the plan, Exponential identifies which capabilities would be useful.
  3. Resolution order — certified → local → community. The first match wins.
  4. User confirmation — for community capabilities, you see a prompt showing which will be used.
  5. Capabilities loaded — the resolved capabilities are made available to the agent during execution.

This means agents gain skills dynamically based on what the task needs — you don't have to manually configure which capabilities each agent has access to.

Writing your own capabilities

You can create custom capabilities for your specific workflows. A capability is just a SKILL.md file:

---
name: My Custom Capability
description: Does a specific thing for my project
tools: [bash, read, write]
tags: [custom, internal]
---

## Steps

1. First, do this...
2. Then, do that...

## Examples

Here's an example of the expected output...

See the Managing Capabilities guide for a full walkthrough, and the Capability File Format reference for the complete specification.

Capabilities vs. agent knowledge

Capabilities and agent knowledge are different:

Agent KnowledgeCapabilities
ScopePermanent — part of the agent's identityPer-task — loaded when needed
PurposeDomain expertise and thinking patternsSpecific procedures and tools
Example"Expert in TypeScript and React""How to export a PDF from HTML"
PersistenceAlways activeOnly active when resolved

Think of knowledge as what the agent knows inherently, and capabilities as what it learns to do for a specific task.

On this page