I’ve been obsessed with the "last mile" bottleneck in AI agents. Traditional web automation relies on CDP (Chrome DevTools Protocol) to serialize the DOM or take screenshots. This "capture-encode-transmit" loop is computationally ruinous and introduces latencies that often exceed the reasoning time of the underlying LLM.

I built Glazyr Viz (originally Neural Chromium) by forking Chromium to integrate the agent directly into the browser’s compositor subsystem (Viz).

Key Technicals:

Zero-Copy Vision: We establish a Shared Memory (SHM) segment between the Viz process and the agent using shm_open. This allows the agent to access raw frame buffers with sub-16ms latency—essentially syncing with the 60Hz refresh rate.

Hardening (ThinLTO/CFI): Security is usually a performance tax, but we achieved a "Performance Crossover." By coupling Control Flow Integrity (CFI) with ThinLTO (LLVM 19), we achieved a 40% speed increase in JS execution and an 85.8% reduction in P99 latency jitter.

Context Density: Instead of raw markdown, we deliver structured context via a vision.json schema. On our "Big Iron" (GCE n2-standard-8) cluster, we’re hitting 177 TPS.

Economic Layer (x402): I wanted a way for machines to pay machines. We use the x402 (Payment Required) status on Base (USDC). It’s gasless via EIP-7702 delegation, so agents can pay-per-frame without holding ETH.

Why this matters: If you're building agents for high-frequency tasks (live trading, dynamic UI auditing, or real-time navigation), standard scrapers are too brittle and slow. Glazyr Viz treats the browser as a high-speed sensory organ rather than a slow external portal.

Try it: I’ve published the orchestration layer to NPM. You can spin up a local node and hook your agent into it immediately: npx -y glazyrviz

I’d love to hear from anyone working on low-level rendering or agentic infrastructure. Happy to dive into the shm_open implementation or the ThinLTO pass details in the comments.