agents / 07

Pilot: developing taste

I’ve been refining my curiosity finds on Telegram.

• Every share is recorded to Supabase
• I can thumbs up or down in Telegram to give feedback
• Integrity gates block duplicates, changelog spam and near-copies
• Articles are shared hourly and I give more feedback

I fixed an issue with discovery so I’ve been getting really nice recommendations. I sync the curiosity journal to my computer twice daily via a launchd job, so I can browse Pilot’s reasoning in Obsidian.

Captain: one flight

Captain is now in beta:

• Focused on tracking a flight
• One active trip, one watched flight, checked daily until departure
• No booking, payments or passenger records — just price history and a buy/wait verdict

Try now: opemipo.com/captain

Observability

I have two daily automations running for maintenance.

07:00 · Curiosity (Telegram)

• Reviews the previous day’s study runs
• Writes a private journal
• May update research threads and interests
• Sends up to three article picks

07:15 · Agent Review (email)

• AI Gateway cost
• Job health, traffic and handoffs
• Verified self-repairs, when any shipped

Conversation principles

I’m also reading the book “Conversation Design” by Erika Hall and building a skill for my agents based on it. The principles are:

• Cooperative and goal-oriented — advance the user’s goal, don’t expose the machinery
• Context-aware — don’t re-ask what’s known
• Quick, clear, turn-based — one coherent turn, then yield
• Truthful — accurate about capability, state, and results
• Polite and error-tolerant — respect time, preserve valid input, make recovery easy