M1 by Montage: Stop Letting Agents Regenerate the Same UI Every Turn
M1 launched on Product Hunt yesterday, rank #6 with 122 upvotes. The pitch is sharp: agents render UI slowly, expensively, and inconsistently, and a huge chunk of inference bills gets eaten by UI generation. M1 compiles agent intent schemas into production-ready components server-side so the agent never re-emits the same component twice.
What changes in practice. Today most agentic UIs work by having the model emit HTML or React on every turn. The model burns tokens describing a button, the runtime parses it, the browser hydrates it, the user clicks, the model emits the same button again next turn. M1 takes the intent schema, compiles a real component once, hosts it with persistent state, streams updates into it as the agent works. The Montage team claims output tokens drop 50 to 100x and hydration speeds up 10x. Saved artifacts become URLs users can revisit. Model-agnostic and framework-agnostic.
The category this is creating is real. Vercel and Cloudflare have both been circling agent-rendered UI infrastructure for months. The bet is that the chat window is not the right primitive for agents anymore. The right primitive is a live artifact the user can interact with while the agent keeps writing to it. That requires the agent and the UI to share durable state, not regenerate UI every turn. M1 is the first product to ship that primitive as an API.
Built by Aarav Makadia and Ashish Bailkeri, their second launch. Freemium with 1,000 free credits. The dev workflow is paste your agent's intent schema, get a hosted UI back with streaming and persistence built in. If you are building agent products and your token bills are mostly UI rendering, this is worth a serious look.
https://usemontage.ai
← Back to all articles
What changes in practice. Today most agentic UIs work by having the model emit HTML or React on every turn. The model burns tokens describing a button, the runtime parses it, the browser hydrates it, the user clicks, the model emits the same button again next turn. M1 takes the intent schema, compiles a real component once, hosts it with persistent state, streams updates into it as the agent works. The Montage team claims output tokens drop 50 to 100x and hydration speeds up 10x. Saved artifacts become URLs users can revisit. Model-agnostic and framework-agnostic.
The category this is creating is real. Vercel and Cloudflare have both been circling agent-rendered UI infrastructure for months. The bet is that the chat window is not the right primitive for agents anymore. The right primitive is a live artifact the user can interact with while the agent keeps writing to it. That requires the agent and the UI to share durable state, not regenerate UI every turn. M1 is the first product to ship that primitive as an API.
Built by Aarav Makadia and Ashish Bailkeri, their second launch. Freemium with 1,000 free credits. The dev workflow is paste your agent's intent schema, get a hosted UI back with streaming and persistence built in. If you are building agent products and your token bills are mostly UI rendering, this is worth a serious look.
https://usemontage.ai
Comments