The Capability Spectrum
Rule-based
No memory
Helpful
Limited actions
Tool use
Autonomous
There's a lot of "AI agent" marketing that doesn't mean much. Most apps labelled as AI agents are sophisticated chatbots with a nice UI. Understanding the real difference helps you make better architecture decisions and set honest expectations with users.
What a chatbot actually is
A chatbot takes an input, processes it through a fixed flow, and returns an output. Even an LLM-powered chatbot is still fundamentally reactive — it waits for input, produces a response, and that's the end of the interaction.
Chatbots are great for: answering questions, generating content on demand, helping users draft things, providing information.
What makes something an actual agent
An agent has autonomy — it can take a goal and work toward it without hand-holding. The key capabilities:
- Tool use: calling external APIs, reading files, querying databases, sending messages
- Multi-step planning: breaking a goal into a sequence of actions
- Observation: seeing the results of its actions and adjusting
- Persistence: continuing to work toward a goal across multiple steps
A chatbot helps you write a grocery list. An agent goes and orders the groceries.
The practical middle ground
Most production apps don't need a fully autonomous agent. They need agentic features — specific workflows that benefit from multi-step AI reasoning.
Example: in Roboto Cart AI, the "plan my week's shopping" feature is agentic. It reasons about your pantry, checks meal plans, considers dietary restrictions, and generates an optimised list. The rest of the app is a sophisticated assistant.
When to use a chatbot vs. an agent
Use a chatbot when…
- Answering questions from a knowledge base
- Helping users draft text
- Providing single-step recommendations
- Simple Q&A flows
Use an agent when…
- The task requires multiple steps
- External actions need to be taken
- Context needs to persist across a workflow
- Decisions need to be made mid-task
The honest conversation about reliability
Agents fail more often than chatbots. Multi-step reasoning means more opportunities for the model to go off track. Production agents need error recovery, logging, and human-in-the-loop checkpoints for high-stakes decisions.
Don't build an agent because it's cool. Build an agent when the task genuinely requires it, and you've done the work to make it reliable.