Over roughly the last two weeks we’ve been shipping across the whole product in a way that only shows up as “small fixes” until you step back and see the pattern: we’re making Agentwood behave like a serious consumer platform, not a brittle demo.

On the chat and AI side, we tightened how models are chosen and retried, moved defaults toward current Gemini generations, and improved what happens when the provider is overloaded, especially in streaming flows, so users see clearer failures instead of conversations that mysteriously stall. We also hardened core paths so chat doesn’t collapse just because an optional retention-related database isn’t present in a given environment. That kind of work isn’t flashy on a changelog; it’s what keeps threads trustworthy at scale.

Voice got meaningful attention in the same window: playback behavior on mobile and Safari, scenario audio on touch devices, stability fixes around TTS loops and incorrect seeds, and continued evolution of how synthesis is routed so production isn’t locked into a single fragile engine path.

On surface area and conversion, we refreshed the home experience with a stronger hero and cleaner shell, shipped /vs and broader SEO-driven content work, and pushed forward video replies and dramas alongside landing updates so discovery matches what the product actually does now. Onboarding moved toward something closer to a guided guide / book-style experience, and we introduced a Wood-first-purchase prompt that triggers after real engagement, capped daily so it stays fair rather than spammy.

We also invested in the parts users feel as identity and polish: human/live lobby flows improved, chat bubbles better reflect who is speaking and which thread you’re in, and live-chat presentation fixes matter because “anonymous gray boxes” quietly kills retention even when the model is strong.

On growth instrumentation, we implemented Reddit pixel plus server-side conversion plumbing, including purchase signals reconciled with Stripe, because scaling acquisition without honest attribution is how teams lie to themselves. In parallel we expanded Wood funnel tracking and kept journal-style SEO publishing moving with sitemap updates—compound discovery work, not one viral moment.

Finally—and this is the unglamorous layer that determines whether you survive spikes—we added health and readiness signals for monitors, improved uptime-facing behavior around model availability, enabled row-level security on publicly exposed database surfaces where appropriate for our Supabase-shaped posture, and fixed automated daily content execution so it runs from GitHub Actions at the repo root, where Actions actually loads workflows.

Put together, the thesis is simple: character AI as a category has demand (you can feel that in the signup curve—north of 100,000 accounts is a real signal). What separates products that turn demand into a durable business from products that stall is weekly compounding across model reliability, UX clarity, monetization loops, measurement, and backend discipline. That’s the setup that makes extreme upside plausible—not because we’re promising a timeline, but because we’re building the kind of system that doesn’t break every time traffic doubles.

We’re still early in that arc. The last two weeks weren’t fireworks; they were foundation—and foundation is what makes the next chapters inevitable instead of accidental.