Replit HQ Livestream: Agent 4 Launch Part 1
Key point
Right after the Agent 4 launch, a Replit HQ livestream showcased collaborative building and real demos.
Details
Wrapping up Agent 4 launch week, Manny Bernabe and Raymmar Tirado opened the doors of Replit HQ and livestreamed the team's atmosphere and build process for 2 hours. The core message: thanks to Infinite Canvas, parallel work, and unified building, more people can now collaborate more easily within a single project.
Manny Bernabe built Taste, an app that learns taste and design sensibility, using Agent 4. When you feed it an image or screenshot, Google Gemini analyzes typography, color palette, layout, and visual hierarchy, then turns that into a style prompt that can be pasted directly into other Replit builds.
His project had three outputs together on a single canvas.
- Landing page
- Web app
- Native mobile version
A photo taken with his phone entered the web library within seconds, and the web and mobile versions ran on the same backend. He also generated three landing page variants based on an initial inspiration image: warm and intimate, dense and formal, and playful and electric.
Raymmar Tirado presented Replitopolis, a 3D city built by connecting a BigQuery data lake in read-only mode. Each building represents a real user, and its height reflects the number of prompts or checkpoint commits. When a commit happens, an orange light turns on, vehicles move following traffic rules, and the whole city updates in real time.
During the demo, he also showed version control in action, rolling the project back to its state 4 days ago and then cleanly restoring it, explaining how practical version management works within Replit. At the time of the stream, the #1 spot on the leaderboard had recorded 1,100 prompts and 232 checkpoints in a day.
The message was clear: if company data is read-only, you can safely build internal tools or visualization apps through connectors like BigQuery, Databricks, and Snowflake. In other words, you can produce genuinely useful results without writing to the database.
Tala Awwad from the marketing team announced that recruitment for the creator program was open, and introduced replit.com/creators as the place to apply. The Aura Globe shown on screen during the stream illustrated the global community's expansion, reaching Romania, Dubai, and Melbourne.
Zade Keylani revealed that Agent 4's design was made using Agent 4 itself. Instead of a Figma file, the engineering handoff was a Replit prototype, with every state, screen, and admin panel built inside Replit. This allowed the team to find and fix real problems while refining the product at the same time.
The change he emphasized was that Infinite Canvas lowers the cost of failure. Previously, there was worry that if the agent changed something, it might ruin existing work, but now you can generate multiple ideas simultaneously, keep only the ones you like, and branch off into a new one if needed. It's become more natural for designers, writers, operations staff, and tech leads to each grow the same project together in their own threads.
Raina Saboo explained that the marketing team, which had 2 people during Agent 3, grew to 5-6 people for Agent 4, and DevRel also expanded from 1 person. They had to persuade first-time users, power users, and enterprise customers simultaneously, and a funding announcement and brand refresh overlapped in the same week.
The framework that tied all of this together was low floor, high ceiling.
- Anyone should be able to get started easily
- At the same time, anyone should be able to build anything
The night before launch, the enterprise and marketing teams built complete demos from start to finish themselves, and early access customers provided quotes, demos, and logo approvals at a pace that normally takes months. The conclusion was one: Agent 4 wasn't designed to just increase autonomy, but as a creativity-centered tool that amplifies human intent even more.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.