Studio2025–presentSolo founder · Claude collaborator

AstroSense

Signals-first analytics for agent runs. Trace quality, cost, and failure modes across multi-step agent executions, with dashboards tuned for operators rather than slide decks.

The observability layer agent teams actually use — built from the on-call posture, not the demo posture. Same instinct as Agent Parts, applied to runtime instead of build time.

AstroSense is signals-first analytics for production AI agents. Trace quality, token cost, latency, failure modes — surfaced in dashboards designed for the operator on a 3am page, not the executive on a quarterly review.

Currently in active build. The data plane is built; the dashboard surface is the work in progress, and that's the half that decides whether a tool gets used.

What it is

A small set of operator-shaped views over agent execution data:

  • Trace quality. Per-run pass/fail summary plus the actual trace, not a flat log dump.
  • Cost surface. Token spend by model, by tool, by user — the dimensions an operator slices on when something looks wrong.
  • Failure modes. Cluster of similar failures over time, surfaced with examples — so you fix the cluster, not the most recent ticket.
  • Comparison view. Today's run vs. yesterday's run, this week vs. last, with the same trace shape side by side.

Why I built it

Every AI ops dashboard I've used at client sites had the same problem: it was a thin shell over the trace JSON. Operators got a list of fields; they had to translate the field names into the question they actually had, then squint at the data to answer it.

AstroSense flips that. Start from the questions an operator asks at 3am — Is this run worse than yesterday? What changed? Where's the spend going? — and design the surface to answer them in two clicks.

How it ships

The pattern is the same as the rest of the studio: built with Claude as a daily collaborator, observable from day one (yes, AstroSense observes itself), opinionated about what makes the cut into the operator view.

Stack: React + Edge runtime for the dashboard, an ingestion layer that accepts traces from any agent framework with a small adapter shim. Agent Parts emits compatible traces by default.

Why this matters for AI systems

Observability is where most AI teams discover that "we'll add it later" was the wrong call. By the time an incident hits production, the trace shape is wrong, the labels are missing, and the dashboard is being designed in the same hour you're trying to debug.

AstroSense is the version where the dashboard already exists when the first agent ships. The operator surface is part of v1, not part of the post-mortem.


Notes for verification (TODO before publishing): confirm whether "data plane built, dashboard in progress" matches the actual status. Replace the framing on what's shipped vs. what's coming if I have it inverted. Real launch date, beta sign-up URL, screenshot when available.

ObservabilityLLM PipelinesReactEdge

Working on something similar?

Book an intro call and I’ll walk you through how the same systems thinking applies to your AI work.

Book an intro call