Agent Studio

The Problem

Sketching how a few agents should hand off — intake, planner, orchestrator, executor — usually happens in your head or on a whiteboard, then gets lost in code. There’s no quick way to draw the graph, run it against a real model, and see where it breaks before committing to an implementation.

What I Built

Agent Studio canvas — an Intake → Plan → Execute agent graph with connected nodes, a node inspector (title, description, role, instructions) on the right, snapshots, and a test query run against a local Ollama model ::border

A local visual canvas for designing and testing small agent graphs against a locally-running Ollama instance — no cloud. You drag nodes (agent, tool, subagent roles), connect their ports into a DAG, then run the graph with live per-node status streaming. Each run writes a transcript.json and a brief.md.

Features

Everything runs locally; state lives in localStorage with a panel to promote it to disk. Beyond the basic run loop:

Per-node mocks and fixtures — pin a node’s output to test the rest of the graph deterministically
Solo-node and step runs — execute one node, or step through with an inspector
Snapshots — save and restore canvas states
Import / export — round-trip a graph as Markdown or a JSON spec

Agent Studio onboarding — the three core gestures (drag a node, drag from a port to connect, click a node to edit) over the live canvas ::border

The whole point is a fast loop between sketching an agent topology and watching it actually run against a model you control.