agents / 06

Self improvement

Both Captain and Pilot now have repair loops.

Every day, a test checks whether each agent is still doing its job. If something breaks, another model digs into what went wrong, proposes a fix, and waits for me to approve it. Nothing goes live on its own. Once a fix lands, Pilot’s morning note tells me what changed.

This sits alongside whatever I change from using them myself.

Agent goals

I’ve narrowed the agents down to something useful.

Pilot runs a curiosity study: it reads my bookmarks and a fixed set of feeds for things that match my interests.

Captain tracks upcoming flights and reports the best options and price changes.

Pilot: Curiosity study

Curiosity study is Pilot’s reading loop. Every two hours or so, it goes looking for something worth paying attention to.

I keep a public Raindrop, which makes it easy for Pilot to track my interests and pull RSS where it exists. The core sources include Aeon, The Marginalian, kottke, Dense Discovery, Robin Sloan, Apartamento, and my own Substack. Rolodex expand runs weekly and can add new feeds when they match my interests.

It also reads my website. When it digests one of my essays, it updates its notes on how I write and what I care about.

Separately, it can ping me on Telegram when it finds something good. A few messages a day at most, and only when something is actually worth sharing.

Captain: Price tracking

For now, Captain tracks flight prices. It doesn’t book. Inventory is still the hard part, so I’m keeping the product focused on watching routes well before I trust it to recommend a purchase.

I originally set up a Codex worker to find results, and although it worked, it was a bit hacky. Once I decided to make Captain open, that system stopped making sense. I tried OpenAI web search against a curated list of sites, but that never got reliable results. So I decided to focus on the tracking experience first.

What next

Over the next few days, I’m going to keep testing both agents and see what the repair loops actually ship. Then I’ll open Captain up for anyone interested to test along with me.