Faster Chat, Cheaper Runs: The Harness Speed Release

The chat spinner is dead. Instead of asking a question and staring at a pulsing dot while the whole answer assembled off-screen, you now watch the reply arrive token by token, the moment the assistant starts thinking. That single change — end-to-end streaming into the chat bubble — is the difference between waiting on software and having a conversation with it.

Behind that new responsiveness is a top-to-bottom tune-up of the engine that runs the assistant. We started by wiring up timing instrumentation across every step of a chat turn, which turned vague slowness into a map of exactly where the seconds were going. Then we fixed the offenders: bulk database reads replaced dozens of small round-trips, and the pieces of each request that stay the same got cached instead of rebuilt every turn. The result you feel is blunt — a first reply that used to take around 21 seconds now lands in about 7, and follow-ups dropped from 17 seconds to roughly 5.

We also put the assistant on a diet. A careful review of what chat actually needs revealed it was hauling around a huge pile of tools it never used, so we trimmed it down to what belongs: reading, understanding, and advising. Every action that changes something — planning an intent, editing a brief, wiring an automation — already has a clear, dedicated home in the interface, so chat now focuses on being fast and sharp rather than doing everything twice. Trimming that baggage cut the size of each request by nearly 90%, and moving read-only chat onto a lighter, quicker model dropped the cost of a warm turn by well over 90% while keeping answers just as good. That's not just our bill — it's snappier replies for you.

A few smaller things round out the release. We caught a bug where a brand-new idea would briefly flash as "Ready to build" before it had actually been scoped — now the label waits until there's a real plan behind it. Credit history got friendlier, familiar labels so your transaction log reads the same as it always has. And we squashed a backend gremlin where an automated analysis could fail outright by reaching for a tool it wasn't allowed to use; it now only ever sees the tools it can actually run, so those runs finish cleanly. We also cleaned up some noisy duplicate logging on our side while we were in there.

The upshot: an assistant that answers in seconds, streams as it goes, costs a fraction of what it did, and quietly stopped tripping over itself. Ask it something and watch it work.