Three different shapes
Copilot, assistant, and agent point at different autonomy levels. The terms get used interchangeably; the products that win are clear about which one they're shipping.
Copilot
- Suggests inline; user accepts or ignores.
- Low risk, low surprise. Works because the user is always in the loop.
- Examples: code completion, smart compose.
Assistant
- Acts on explicit request, often with a single confirmation.
- Tool layer required — MCP, function calling, or your own.
- Examples: ⌘K action bars, in-product chat that actually does things.
Agent
- Pursues a multi-step goal with limited human input.
- Needs robust audit, dry-run modes, and rollback.
- Examples: background research, scheduled cleanup, batch processing.
Sequencing
Ship assistant first. Copilots feel passive; agents demand trust you haven't earned. Assistants prove the value of your tool layer and pave the road to both.
FAQ
- Is this just marketing wordplay?
- Partly — but the three terms point at different autonomy levels and product shapes. Getting the term right calibrates what users expect and what you need to build under it.
- Definitions in one line each?
- Copilot: suggests, you accept. Assistant: acts on request, you confirm. Agent: pursues a goal across many steps, you supervise.
- Which should I ship first?
- Almost always assistant. Copilots feel passive; agents demand trust you haven't earned. Assistants prove the value of your tool layer before you scale autonomy.
- When does an agent make sense?
- When the task is multi-step, repetitive, and the cost of supervision is greater than the cost of a bad call. Background research, scheduled cleanup, batch processing.