Core Concepts

This page gives you the full picture. Every concept in Exponential, explained briefly. Dive into individual pages for depth.

The mental model

You (team lead)
 └── Space (workspace)
      └── Task (unit of work)
           └── Agent (AI teammate)
                ├── Personality — who the agent is
                ├── Knowledge — what the agent knows
                ├── Skills — what the agent can do
                └── Provider — where the agent runs (Claude, Gemini, etc.)

You create tasks. You assign agents. Agents plan, you approve, they execute, you review. That's the loop.

Tasks

A task is a unit of work. It has a title, description, and status. Tasks live in a space and move through a lifecycle:

todo → planning → working → review → done

Tasks can also be cancelled or error if something goes wrong. See Task Lifecycle for the full breakdown.

Agents

An agent is an AI teammate with a defined identity. Each agent has:

Personality — how it thinks and communicates
Knowledge — what domain expertise it brings
Principles — what rules and standards it follows

Agents come in three tiers: built-in (pre-configured), community (shared by other users), and custom (written by you). See Agents.

Skills

A skill is a reusable capability. Think of skills as tools an agent can pick up — PDF export, code review, web research, slide creation.

Skills are resolved automatically based on what the task needs. They follow a trust hierarchy: certified (verified by Anthropic) → local (your custom skills) → community (shared by others). See Skills.

Spaces

A space is a workspace container. Each space has its own tasks, agents, memory, and context. You might have one space per project, per client, or per area of work.

Spaces also have memory — persistent context that agents can access across tasks. Add notes like "we use TypeScript strict mode" or "the API follows REST conventions" and every agent in that space will know. See Spaces & Memory.

Providers

A provider is the AI runtime that powers an agent. Exponential supports multiple providers:

Claude (Anthropic) — the default, most capable option
Gemini (Google) — alternative runtime
Codex (OpenAI) — alternative runtime
Copilot (Microsoft) — alternative runtime
Qwen (Alibaba) — alternative runtime

Each agent can use a different provider. See Providers.

The plan-first model

This is the core of Exponential's workflow. Before an agent makes any changes, it proposes a plan. You review, approve, or redirect. Only after approval does the agent execute.

This is intentional. Trust is built through transparency. You always know what's about to happen before it happens. See Task Lifecycle for how this works in practice.