Skip to Content
StudioOverview

Loomflow Studio

Studio records what your agents do so you can look at it afterwards: every model call with the messages sent and the reply, every tool call with its arguments, result and permission decision, every guardrail hit, sub-agent and workflow step, with tokens, cost and time. You turn it on once; nothing about your agents changes, and nothing leaves your machine. Runs go into one SQLite file, and loom studio opens a web page over it.

Quickstart

Turn recording on, either in code:

import loomflow.studio
 
loomflow.studio.enable()

or with an environment variable, which needs no code change at all:

export LOOMFLOW_STUDIO=1

Run your app as usual. Every Agent and Workflow run in the process is recorded to .loomflow/studio.db in the current folder. Then browse it:

pip install 'loomflow[studio]'
loom studio                      # http://127.0.0.1:7600

Run loom studio from the same folder, or pass --db with the file’s path. Recording itself needs nothing beyond the standard library; the studio extra only adds the web server.

Get started walks through it step by step: where to turn it on in a script, a web app or a notebook, naming agents, opening the page, and what to do with it day to day.

What you see

  • Traces. One row per run: status, agent and version, input, user, model, tokens, cost, time and score. Filter by agent, user, model or status, show only errors and interrupted runs, or search everything recorded (prompts, replies, tool arguments and results). Click a session_id to see a whole conversation.
  • A trace. The run as a timeline: the agent run, its turns, model calls, tool calls, guardrail hits, and any sub-agent or workflow steps, nested the way they ran. Click a step to see what it did: the messages a model call sent and its reply, a tool call’s arguments, result and whether it was approved, a guardrail’s verdict.
  • Agents. Every agent and workflow that has run, with its versions. Studio records a new version whenever the instructions, model, tools, team or settings change, and notes what changed, so a regression points at the change that caused it.
  • Architecture. Each version drawn as a diagram: the sub-agents in a team and how they delegate or hand off, a workflow’s steps, routes and parallel branches, and the tools each agent holds. On a trace, the same diagram highlights the agents, steps and tools that run used, so you see which path a request took.
  • Experiments. Each EvalHarness run, grouped by dataset. Compare two runs case by case: which cases improved, which regressed, and how cost and latency moved.
  • Scores and datasets. Mark a run good or bad on its page, and add a bad one to a dataset with the answer it should have given. Then run your evals on that dataset with loomflow.studio.dataset("name").

Try it on a demo

examples/34_studio.py records a small support bot with scripted models, so it runs offline with no API key: two versions of a billing agent, a tech-support agent, a routing workflow, a supervisor team, an error, a run that hits its turn limit, and two eval runs to compare.

python examples/34_studio.py
loom studio --db .loomflow/studio-demo.db

Studio is a local tool: one file, no account, no server to run besides the page itself. It records loomflow runs. For monitoring production across a fleet, also export OpenTelemetry to your tracing backend; the two work side by side.

Next

Last updated on
AgentGraph visualizationGet started