Agent Studio

Agent Studio is where you create agents: name, personality, models, tools, and (optionally) a swarm of child agents.

Open Agents in the sidebar, or press Ctrl + K and type Create Agent.

---

Create an agent

1. Click + New Agent.
2. Fill in identity:
   • Name — what you see in Chat (for example Research Assistant).
   • Slug — stable id (for example research-bot).
   • Description — a short role, useful for you and for other agents in a swarm.
3. Write a system prompt that states the job and the limits.
4. Choose models, tools, and approval level.
5. Click Save Agent Manifest.

A practical prompt looks like this:

# Role
You are a research assistant on ActonOS.
Find sources, check facts, and write clear Markdown summaries.

# Limits
1. Prefer Workspace files the user named (for example go_126_summary.md).
2. Ask for approval before restarting services or deleting files.
3. If a search fails, say so instead of guessing.

Prompt variables

You can insert:

• {{system.time}} — current time and timezone
• {{system.runtime_mode}} — baremetal or docker
• {{workspace.root}} — Workspace path

---

Models

ActonOS tries the primary model first. If that model is rate-limited or errors, it falls through the fallbacks.

Setting	What it does	Typical value
Primary model	Everyday reasoning	anthropic:claude-3-7-sonnet
Fallback 1	Used when the primary fails	openai:gpt-4o
Fallback 2	Cheap or local last resort	ollama:qwen2.5-coder
Temperature	0 is strict, 1 is more creative	0.2 for code and analysis
Max tokens	Cap per reply step	4096

---

Tools and safety

Give each agent only the tools it needs:

• * — every installed tool
• native_* — built-in system tools only
• mcp_github_* — GitHub MCP tools only

Approval level:

Level	Behavior
Low	Authorized tools run without asking.
Medium	Workspace writes can run. Shell, delete, cron, and extension changes still need you.
High	Confirm every sensitive action.

Also set:

• Allowed Workspace paths — for example docs/* and reports/*.
• Iteration budget — maximum reasoning steps per task (default 25).

---

Swarms (optional)

A parent agent can hand work to specialists:

1. Open Swarm & Delegation.
2. Enable Agent Delegation.
3. Pick allowed child agents (for example Researcher and Code Writer).
4. Set a delegation budget (how deep nested calls may go, for example 3).

The parent stays in charge of the goal; children run their own sandboxed steps and return results.

---

Related pages

• Chat — talk to the agent you just saved.
• Missions — long jobs with checkpoints.
• Tools & Skills — what tools you can grant.