agents / 08

Testing Captain

The agent is now usable: t.me/captainflightbot

You can talk about travel, watch flight prices, or plan for a trip. It can work through complex itineraries, e.g.

I’m going to Paris in November for a wedding from Nov 4–8 and a birthday in Marseilles on the 14th. Then I have a panel in New York on Dec 10, and I want to spend Christmas in Lagos. What’s the cost looking like for max one-stop flights.

You can plan one trip at a time. Send text or voice in Telegram, and use commands for quick actions. There’s also a visual workspace to browse agent activity, set preferences, or change decisions.

I made a website to explain: moonlight.ng/captain

Observability

I’ve also built an admin for tracking agent costs and conversation.

Feels like this is just the beginning of the software and it only gets better (and more expensive) through use. That’s always kind of been the case, but the architecture of LLMs makes it clear.

Pilot’s body

The Raspberry Pi already keeps a cold mirror of Pilot’s code, memory and database. Now it also has a small e-paper status screen.

• Refreshes every five minutes: time, Pi capacity, memory mirror freshness
• Day’s calls, tokens, cost and failures across Pilot, Curiosity and Concierge
• Never displays memory, prompts or conversations
• Sleeps after each refresh; the image stays visible

I’m also designing/printing a case.