Studio Agent
Watch your agent think.
Overview
Four pillars: orchestration, context engineering, observability, and evals. Every tool call is persisted as a run step — which tool, what went in, what came out, how long it took — so "why did it say that?" is a query rather than a reconstruction.
The problem
Without traces, agent failures are reconstructions. Eval scores can lie when the judge's training data predates the evidence in the prompt.
Wins
The first eval run averaged 2.4 out of 5 until traces proved the judge was wrong about real search results. Observability paid for itself the first time the evals lied.
Tech stack
FastAPI / LangGraph / Anthropic / Postgres / SQLAlchemy / Next.js 16 / Render