Explainer
AI Agent vs. Chatbot: What's Actually Different
A chatbot responds to prompts: you ask, it answers, the exchange ends. An AI agent pursues a goal across time — it holds persistent memory of your situation, decides its own next steps, and can act or follow up without being prompted again. The practical test is who initiates. If nothing ever happens until you type something, it's a chatbot, whatever the marketing says.
The distinction, concretely
| Chatbot | Agent | |
|---|---|---|
| Who starts | You, every time | Either side |
| Memory | The current conversation | Persistent, cumulative |
| Goal state | None between sessions | Survives and is tracked |
| Next step | You decide | It proposes and tracks |
| On a day you don't open it | Nothing happens | Work continues |
Why most "AI agents" are chatbots
"Agent" became a marketing term considerably faster than it became a technical reality. Putting a chat interface over a model API is a weekend; building something with memory that persists, goal state that outlives a session, and any mechanism for acting unprompted is a different order of problem. The gap between the two is where most of the category currently sits.
The filter that cuts through it: does this product ever do anything on a day I don't open it? If the answer is no, the agency is aspirational.
What real agency requires
- Persistent memory.Context that carries between sessions, so you're not re-explaining your situation weekly.
- Durable goal state. Objectives that exist independently of any conversation and can be checked against reality later.
- The ability to initiate. Some path by which the system reaches you rather than only the reverse.
- Orchestration. Judgment about how to handle a given task — including which model should handle it — rather than routing everything identically.
Any one of these alone isn't enough. Memory without initiation is a chatbot that remembers. Initiation without memory is a notification.
Where Refractive fits
Refractive is built as an agent by this definition. It keeps a persistent Context Graph of what it knows about your life, tracks missions across eight life domains so goal state outlives any single conversation, initiates check-ins rather than waiting to be opened, and routes work to different models by complexity so you never think about cost or model choice. It's an AI harness — model-agnostic, running on open-weight models by default or on your own Anthropic, OpenAI, Google, or OpenRouter key — rather than a model itself.