Agents
Hand any task to Claude Code, Codex, or OpenCode without leaving your tool, and keep the conversation in your project.
What agents give you
Prompt Studio has no coding agent of its own. It runs the agents you already use, through harness extensions. A harness starts the agent, sends it your messages, and turns its output into a conversation that Prompt Studio shows and saves.
Prompt Studio ships harnesses for Claude Code, Codex, OpenCode, and Pi. Choose the ones you want during project onboarding, or turn them on in Extensions. An optional harness extension adds GitHub Copilot. A harness does not install the agent or sign you in; install the agent and sign in with its own tools.
One conversation with an agent is a session. Prompt Studio starts, queues, and stops sessions, saves the conversation, and shows approval requests. A session that finishes only means the agent stopped. Check the result before you rely on it.
Use agents in Prompt Studio
Choose New conversation on the project’s Home page. The conversation opens in a side panel, so the tool you are working in stays visible. Pick an agent and model, type your request, and send it.
- The workspace selector above the message field chooses where the agent works. Project folder uses your project’s own files.
- When an agent asks before running a tool or changing a file, approve or deny the request in the session.
- When too many sessions run at once, new ones wait as
queuedand start in order. In Settings → Workbench → Runtime, turn on Limit concurrent sessions and set Max concurrent sessions; turn it off for no limit.
From the terminal, run these inside the project folder:
pst agents list
pst sessions create --prompt "Summarize this folder" --agent pstdio.harness-claude-code.harness.claude-code
pst sessions follow-up --id <session-id> --prompt "Now add a README"
pst agents list shows each harness, its full ID, and whether its agent is installed. If an agent is unavailable, the list and the harness menu say why, for example a version that is too old.
Install your tools’ skills for an agent with pst agents setup <agent-id>, so it knows how to use them. See Skills.
Use agents from your tool
A tool can start and continue agent work through ctx.sessions. ctx.sessions.create({ title, prompt, workspaceId }) starts a session, ctx.sessions.followup sends another message, and ctx.sessions.query lists the project’s sessions. A hook on sessionEvents.completed lets the tool react when the agent finishes. Planner uses the same API to run agent attempts on its tickets.
To connect another agent, write a harness extension and declare it with defineHarness in the harnesses contribution. The Codex harness source is a complete example.
Learn more
- Run agents installs an agent and starts a session.
- Agents and harnesses explains what the host and the harness each do.
- CLI sessions lists every session command.
- Commands and processes describes
ctx.sessions. - Codex harness source shows a full harness.
Updated at: 2026-11-22