Falcon AI and 4x Faster Frontend

Falcon AI: An AI Assistant That Knows Your Context

Most AI assistants embedded in SaaS products work the same way: they sit in a sidebar, ask what you need, and reply in a generic voice. The promise is speed; the reality is usually an extra tab with no idea what you were doing.

Falcon takes the opposite approach. It reads the page you are on. It already knows which trace, evaluation, dataset, or simulation you are looking at, so when you ask a question, it starts from the object in front of you, not from a blank slate.

What’s new

• Debug traces from inside the trace. Open a trace where the output drifted and ask Falcon to debug it. It walks the span tree, finds the step where the model went sideways, and explains what likely went wrong, using the actual span input and output, not a generic guess.
• Scaffold simulations in one prompt. Ask Falcon for a simulation to test a fix and it pre-fills the setup from the trace you are inspecting and starts the run.
• Describe, don’t build.
• Describe an evaluation in plain English → Falcon scaffolds it.
• Hand it a few example inputs → Falcon generates a dataset matching the shape of what you showed it.
• Half-remember a trace from yesterday → describe it and Falcon finds the real one.
• Permissions inherited, never loosened. Falcon reads only what you can read and acts only where you can act. Every action lands in the activity feed under your name.

Why it matters

The cost of using an observability platform is mostly navigation: finding the right trace, evaluation, dataset, and stitching them together in your head. A context-aware assistant does that stitching inside the conversation, collapsing the round trip from “something looks off” to “here is a reproducible test case that catches it” into a single thread.

Who it’s for

Anyone who spends half the day clicking between traces, evaluations, datasets, and simulations trying to follow one thread. And anyone new to the platform who doesn’t yet know where everything lives. Falcon does, and will take you there.

Available today across the platform. Read the docs →

4x Faster Frontend

Four targeted optimizations together delivered a four-times improvement in page load performance across the platform. Not cosmetic changes, actual architectural work:

• Route-level code splitting. Stop loading JavaScript the current page doesn’t need.
• Aggressive data prefetching. API calls start before navigation completes.
• Virtualized rendering for large lists and tables.
• Optimized state management that eliminates unnecessary re-renders.

The pages it matters most on: trace views, evaluation dashboards, and simulation results. Previously 2-3 seconds to interactive; now under a second.

Infrastructure and Reliability

The ClickHouse trace storage introduced two releases ago now runs on Replicated MergeTree, adding automatic replication and failover. If a storage node goes down, queries continue against replicas without interruption, delivering the high-availability guarantee that production-critical observability requires.

Improvements

Prompt generation and improvement. Describe a task in plain English and the platform drafts a starting prompt. Already have a prompt? One-click improvement suggestions analyze it and propose specific changes to improve clarity, reduce ambiguity, and better constrain model behavior.

Voice simulation moves to LiveKit. Voice simulation now runs on LiveKit instead of Vapi, with lower latency and more natural conversational dynamics during simulated calls.

Security hardening. When a user disables a 2FA/TOTP device, the change propagates to all replicas right away, so a revoked device cannot keep authenticated traffic alive. Plus a custom admin view for managing blocked IPs and users.

Grid pagination. All list grids in the platform now load 30 rows at a time and continue with infinite scroll, with the next page prefetched while you scroll. Each row’s payload was trimmed so the grid stays responsive on large datasets.