Conversational assistants

We want agents that talk with us and also do things for us. A conversational assistant is one way in: she holds a thread, chooses actions, and keeps nudging a task along while we stay in plain language. We give her a goal; she handles the back-and-forth.

Following the thread

AI can track a conversation across turns. After the first mention, she keeps the goal in working memory while we push the details around. When we change part of the request, she adjusts without losing the rest. It feels simple, which is the point.

Choosing the next move

An agent decides what to do next. After we state the goal, she checks what she already knows and whether she needs more. If a gap appears, she asks a question; if a tool will help, she calls it. We avoid micromanaging her. She keeps her own checklist.

Looping with tools

AI can call local tools when a step needs real data. After that first tool call, she uses results to steer the next turn. Maybe she asks us to confirm something; maybe she goes straight to another tool. The loop stays tight because her steps are tied to what we said a moment ago.

Managing small mistakes

AI will drift now and then, so we want her to notice when something doesn’t fit. After a mismatch, she can restate her understanding and wait for a correction. This saves us from rewriting the whole request. Small repairs keep the conversation moving.

Why it works

We get agents that stay conversational while they plan. She talks, checks, and acts in short cycles. It’s not magic; it’s a steady rhythm of goals, tools, and questions. Like a teammate who prefers chat over forms.

A coder’s musing

We keep adding features, yet the best ones hide. When an assistant handles the glue work between our sentences, we forget she’s doing it—until we build one and see how much that glue matters.

“One last drink” text adventure game

Text adventures may be old, but they’re still fun to play. We type short commands, read what happens, and try again. The goal here is simple: convince the bartender to serve you one last drink. Start by looking around, examining anything that seems interesting, and talking to the people you meet.

At its core, this is a traditional adventure game driven by a deterministic verb–object parser. Rather than replacing that engine, we introduce a conversational AI agent as a game master. Player input is first interpreted as intent—command, question, or noise—before being mapped into the engine’s limited vocabulary (for example, turning “Ask the bartender for a drink” into TALK BARTENDER). The agent can also respond to questions using the current game state and conversation history, illustrating how natural language interfaces can wrap existing systems without changing their internal logic.