Shipping in public: day 30

Thirty days in. Here is the honest log. Not the highlight reel — the actual state of the build, including the parts that embarrassed us. We said we'd ship in public, and public means the breakages too.

What shipped

The pricing engine, the dealer wedge, and the first agents.

The pricing engine is the spine of it. It ingests sold listings, auction results, and dealer-to-dealer movement, cleans the mess, and resolves any reference down to a band — market price, bid, ask, and a 30-day trend. A month ago this was a spreadsheet and a lot of manual judgement. Now it answers in under a second for the references that matter most, and it's been right often enough that we've started trusting it in real conversations.

The dealer wedge is the smallest useful thing we could put in a dealer's hand: send a reference, get the band back, no app to install. It lives where dealers already work rather than asking them to come to us. That choice mattered more than any feature — three of them now use it weekly without being asked.

The first agents went live doing the unglamorous work: watching for new listings, flagging price moves, and keeping the reference data fresh overnight. Nothing a human couldn't do. Everything a human hated doing.

What broke

More than we would like. That is the point of building in public.

The ingest pipeline choked twice on listings that didn't match any known reference and silently dropped them instead of flagging them — so for a few days the band on one popular model was confidently wrong. Confidently wrong is the worst failure mode a pricing tool has, and we shipped it. We've since added a "we're not sure" state, because a visible gap is safer than an invisible guess.

The dealer wedge sent one duplicate quote in a thread and looked, for a moment, like it was spamming a customer. Small bug, big trust cost. Fixed the same day, but it taught us that in someone else's conversation, you get one strike.

And we underestimated how much of the first month would be plumbing — auth, rate limits, the boring scaffolding nobody screenshots. The demo took a day. Making the demo dependable took the other twenty-nine.

What the agents handled alone

This is the part that surprised us. By the end of the month, the overnight data refresh, the new-listing sweep, and the first pass of anomaly flagging ran without a human in the loop. We'd wake up to a clean dataset and a short list of things worth a person's attention, rather than to the whole job still waiting.

That's the bet behind all of this: the agents don't replace judgement, they clear the ground in front of it. A month of doing the dull parts unattended bought back enough hours to actually build.

What's next

Tighten the "not sure" handling so the engine degrades honestly instead of guessing. Widen the reference coverage past the headline models. And give the agents one more job — drafting the dealer's reply, for a human to approve — so we keep moving up the value chain from data to decisions.

Day 30 done. The log stays open.