Dhaka, Bangladesh.
Building practical AI businesses.

Thinking / Topic guide

A conversation is only the beginning.

Conversational AI interprets and responds. An agentic system can also use tools to take steps toward a goal. The design question is how much freedom a particular task needs.

TawfiqueAhmed.com / Editorial topic guide
Updated · About the site’s owner

Separate a response from an action

Answering a question about opening hours and changing an appointment have different consequences. An action needs a clear permission boundary, the right information, and confirmation that it succeeded. Design these distinctions into the experience instead of leaving the customer to infer what happened from a fluent answer.

Use the simplest workable system

Some tasks fit a predictable sequence; others need more flexible decisions. Anthropic distinguishes predefined workflows from agents that dynamically direct their own processes and tool use. Start with the simpler approach that satisfies the task. More autonomy introduces more decisions to test, observe, and recover from.

Design the exception before the demo

Consider missing information, contradictory requests, unavailable tools, and a person changing their mind. A reliable system explains uncertainty and offers a route forward. A handoff should carry the necessary context so a customer does not have to begin again. Avoid claiming completion until the underlying operation has confirmed it.

Keep a usable record

For consequential actions, record what was requested, what information was used, what the system attempted, and the result. Protect access to those records and keep them only as long as needed. This makes investigation and improvement possible without treating every conversation as material for unrestricted collection.

Evaluate the whole journey

Test whether the system reaches a useful outcome, including the cases where stopping is correct. Measure unnecessary actions, missed handoffs, and recovery after errors. A convincing demonstration is a starting point; operating quality depends on what happens when the conversation is incomplete, ambiguous, or unexpected.

Further reading

Anthropic — Building effective agents