Agents
An agent in Exponential is an AI teammate. Not a chatbot, not a tool — a teammate with a defined identity, expertise, and working style.
Agent anatomy
Every agent has three core sections:
Personality
Who the agent is. This defines its communication style, thinking approach, and character traits.
## Personality
You are a pragmatic senior engineer who values simplicity over cleverness.
You communicate directly — no fluff, no hedging.
When you see over-engineering, you call it out.Knowledge
What the agent knows. This is the domain expertise the agent brings to every task.
## Knowledge
Expert in TypeScript, React, and Node.js.
Deep understanding of REST API design and database modeling.
Familiar with Tailwind CSS and modern frontend architecture.Principles
How the agent works. These are the rules, standards, and non-negotiables the agent follows.
## Principles
- Always write tests for new functionality
- Prefer composition over inheritance
- Never introduce a dependency without justifying it
- Keep functions under 30 linesTogether, these three sections form the agent's system prompt — the instructions that shape its behavior on every task.
The three tiers
Built-in agents
Pre-configured agents that ship with Exponential. They cover common use cases and are ready to use immediately.
Examples: general-purpose developer, documentation writer, code reviewer.
Community agents
Agents shared by other Exponential users. Browse the agent marketplace to discover agents for specific domains — data engineering, DevOps, frontend development, marketing, and more.
Community agents are imported into your workspace and can be customized after import.
Custom agents
Agents you create yourself. Write a .agent file with YAML frontmatter and markdown body sections (Personality, Knowledge, Principles), and Exponential discovers it automatically.
Custom agents can be stored:
- Globally:
~/.claude/agents/*.agent— available across all spaces - Per-project:
your-project/.claude/agents/*.agent— available only in that project's space
See the Custom Agents guide for a full walkthrough.
How agents execute
When you assign an agent to a task, here's what happens under the hood:
- System prompt injection — the agent's Personality + Knowledge + Principles are combined into a system prompt and sent to the AI provider.
- Context loading — the agent receives the task description, space memory, and access to the relevant codebase or files.
- Planning — the agent explores the context and proposes a plan.
- Execution — after approval, the agent executes using its provider's capabilities (code editing, file creation, terminal commands, etc.).
- Communication — throughout the process, the agent posts updates to the task's activity timeline.
Choosing the right agent
Match the agent to the task:
| Task type | What to look for |
|---|---|
| Code features | Agent with relevant language/framework knowledge |
| Bug fixes | Agent with debugging principles and systematic approach |
| Documentation | Agent with writing expertise and clear communication style |
| Code review | Agent with quality standards and attention to detail |
| Research | Agent with analytical thinking and thorough investigation principles |
When in doubt, start with a built-in general-purpose agent. You can always create specialized agents as you learn what works for your workflow.
Agent configuration
Beyond the core sections, agents can be configured with:
- Provider — which AI runtime to use (Claude, Gemini, Codex, etc.)
- Model — which specific model within the provider
- Default skills — skills that are always available to this agent
- Tags — for organization and discoverability
- Color — for visual identification on the board
See the Agent File Format reference for the complete specification.