Agent Studio
Start here

Your First Task in 30 Minutes

Clone Agent Studio, connect one repository and take one deliberately small coding task all the way to a review decision.

This is the shortest useful path through an early-alpha tool. You build the three binaries from source, connect one repository and run a coding-agent CLI you already use — on your workstation or a remote host. There is no installer or hosted account yet.

Open the repo README Understand the concepts
Before you start

0–5 min · Check the Prerequisites

Bring a Git repository you can safely change and a configured Codex, Claude Code or Gemini CLI. Agent Studio organizes those tools; it does not replace their accounts or installation.

  • Install Git, Node.js and the coding-agent CLI you intend to run.
  • Authenticate that CLI and verify it can open in a terminal.
  • Choose a disposable or clean repository and make sure its current work is committed.
Clone and build

5–15 min · Clone, Build the Three Binaries, Open the Studio

Use the repository README as the source of truth for today’s commands. You build three things from source: the Task Server (the central service and system of record), the Agent Studio web app you open in the browser, and the Agent Runner that executes coding-agent CLIs. The project is early alpha, so the website avoids promising a release package that does not exist.

  • Clone github.com/RobertMischke/agent-studio and enter the checkout.
  • Follow the README to build and start the Task Server — the system of record every client talks to over HTTP.
  • Open the Agent Studio web app in the browser at the address the server prints, then add your target repository as a project.
  • Start an Agent Runner (binary agent-runner) on this machine or a remote host so it can execute the coding-agent CLIs.
  • If setup diverges from the README, stop and report the exact command and output; do not guess around credentials or paths.
git clone https://github.com/RobertMischke/agent-studio.gitcd agent-studio# follow the README to build the Task Server, Agent Studio and agent-runner
What early alpha means
Architecture

The Three Binaries You Build

Agent Studio is remote-first: one central Task Server, a browser-based studio and Agent Runners that can live on any host. Understanding the three parts makes the setup — and where each one runs — obvious.

Task Server

The central service and system of record. It owns task state, run leases and review evidence; every client talks to it over HTTP.

Agent Studio

The web app you open in the browser — not a desktop app. It renders the board, task detail, evidence and review over the Task Server API.

Agent Runner

Runs on any host, local or remote, and executes coding-agent CLIs in isolated git worktrees. The binary is agent-runner.

System map

Central state. Execution anywhere.

Control stays central. Execution lives wherever you allow it.
01Client

Agent Studio

See and decide in one browser surface.

HTTP client
02Authority

Task Server

Remember and coordinate every task, run and review decision.

authoritative state
03Execution

Agent Runner

Execute coding-agent CLIs and report from a host you control.

agent-runner
Studiorequest
Task Serverstate + lease
Runnerstatus + evidence
The Studio and every Runner communicate through the Task Server; no browser-to-runner shortcut owns task state.
Shape the task

15–20 min · Write a Task an Agent Can Finish

Create one bounded task with a visible result. A good first task names the repository, desired change, constraints and proof required for acceptance.

Example outcome

Add an empty-state sentence to the project list.

Example boundary

Change only the list component and its existing test.

Example proof

Run the focused test and include a screenshot at mobile width.

  • Outcome: one sentence describing the observable change.
  • Boundary: files or subsystem in scope, plus what must stay untouched.
  • Proof: the smallest relevant check—test, build, screenshot or exact output.
  • Finish line: the result is ready for your review, not silently accepted.
Run and review

20–30 min · Run, Inspect, Decide

Choose the configured agent, start the task and watch its protocol. When it stops, inspect the change as software work rather than trusting a completion message.

Agent Studio Stable
Task detail view for a quality-gate task with protocol, Git evidence, and review context.

Quality-gate task detail from a pinned DEMO-9 task.

  • Read the changed files and confirm the task boundary was respected.
  • Check the recorded test, build or screenshot evidence.
  • Accept only if the outcome and proof match; otherwise reissue with one concrete finding or split the task.
  • Your first success is a reviewed task with a decision—not merely a green agent status.
Troubleshooting

If the First Run Does Not Finish

Failures are useful when their boundary stays visible. Preserve the task protocol and diagnose the earliest concrete failure before starting over.

  • CLI does not start: verify the same command and authentication outside Agent Studio.
  • Repository is wrong: confirm the project path and current Git state.
  • Task drifts: narrow the outcome and explicitly name out-of-scope files.
  • Proof is missing: request the exact command or screenshot as a reissue finding.